/* -*- Mode: C++; tab-width: 2; indent-tabs-mode: nil; c-basic-offset: 2 -*- */ /* vim: set sw=2 ts=2 et tw=79: */ /* This Source Code Form is subject to the terms of the Mozilla Public * License, v. 2.0. If a copy of the MPL was not distributed with this * file, You can obtain one at http://mozilla.org/MPL/2.0/. */ #include "mozilla/DebugOnly.h" #include "mozilla/Likely.h" #include "mozilla/dom/nsCSPService.h" #include "nsError.h" #include "nsHtml5TreeOpExecutor.h" #include "nsScriptLoader.h" #include "nsIContentViewer.h" #include "nsIContentSecurityPolicy.h" #include "nsIDocShellTreeItem.h" #include "nsIDocShell.h" #include "nsIDOMDocument.h" #include "nsIScriptGlobalObject.h" #include "nsIWebShellServices.h" #include "nsContentUtils.h" #include "mozAutoDocUpdate.h" #include "nsNetUtil.h" #include "nsHtml5Parser.h" #include "nsHtml5Tokenizer.h" #include "nsHtml5TreeBuilder.h" #include "nsHtml5StreamParser.h" #include "mozilla/css/Loader.h" #include "GeckoProfiler.h" #include "nsIScriptError.h" #include "nsIScriptContext.h" #include "mozilla/Preferences.h" #include "nsIHTMLDocument.h" #include "nsIViewSourceChannel.h" #include "xpcpublic.h" using namespace mozilla; NS_INTERFACE_TABLE_HEAD_CYCLE_COLLECTION_INHERITED(nsHtml5TreeOpExecutor) NS_INTERFACE_TABLE_INHERITED(nsHtml5TreeOpExecutor, nsIContentSink) NS_INTERFACE_TABLE_TAIL_INHERITING(nsHtml5DocumentBuilder) NS_IMPL_ADDREF_INHERITED(nsHtml5TreeOpExecutor, nsContentSink) NS_IMPL_RELEASE_INHERITED(nsHtml5TreeOpExecutor, nsContentSink) class nsHtml5ExecutorReflusher : public Runnable { private: RefPtr mExecutor; public: explicit nsHtml5ExecutorReflusher(nsHtml5TreeOpExecutor* aExecutor) : mExecutor(aExecutor) {} NS_IMETHOD Run() override { mExecutor->RunFlushLoop(); return NS_OK; } }; static mozilla::LinkedList* gBackgroundFlushList = nullptr; static nsITimer* gFlushTimer = nullptr; nsHtml5TreeOpExecutor::nsHtml5TreeOpExecutor() : nsHtml5DocumentBuilder(false) , mPreloadedURLs(23) // Mean # of preloadable resources per page on dmoz , mSpeculationReferrerPolicy(mozilla::net::RP_Default) { // zeroing operator new for everything else } nsHtml5TreeOpExecutor::~nsHtml5TreeOpExecutor() { if (gBackgroundFlushList && isInList()) { mOpQueue.Clear(); removeFrom(*gBackgroundFlushList); if (gBackgroundFlushList->isEmpty()) { delete gBackgroundFlushList; gBackgroundFlushList = nullptr; if (gFlushTimer) { gFlushTimer->Cancel(); NS_RELEASE(gFlushTimer); } } } NS_ASSERTION(mOpQueue.IsEmpty(), "Somehow there's stuff in the op queue."); } // nsIContentSink NS_IMETHODIMP nsHtml5TreeOpExecutor::WillParse() { NS_NOTREACHED("No one should call this"); return NS_ERROR_NOT_IMPLEMENTED; } NS_IMETHODIMP nsHtml5TreeOpExecutor::WillBuildModel(nsDTDMode aDTDMode) { mDocument->AddObserver(this); WillBuildModelImpl(); GetDocument()->BeginLoad(); if (mDocShell && !GetDocument()->GetWindow() && !IsExternalViewSource()) { // Not loading as data but script global object not ready return MarkAsBroken(NS_ERROR_DOM_INVALID_STATE_ERR); } return NS_OK; } // This is called when the tree construction has ended NS_IMETHODIMP nsHtml5TreeOpExecutor::DidBuildModel(bool aTerminated) { if (!aTerminated) { // This is needed to avoid unblocking loads too many times on one hand // and on the other hand to avoid destroying the frame constructor from // within an update batch. See bug 537683. EndDocUpdate(); // If the above caused a call to nsIParser::Terminate(), let that call // win. if (!mParser) { return NS_OK; } } if (mRunsToCompletion) { return NS_OK; } GetParser()->DropStreamParser(); // This comes from nsXMLContentSink and nsHTMLContentSink // If this parser has been marked as broken, treat the end of parse as // forced termination. DidBuildModelImpl(aTerminated || NS_FAILED(IsBroken())); if (!mLayoutStarted) { // We never saw the body, and layout never got started. Force // layout *now*, to get an initial reflow. // NOTE: only force the layout if we are NOT destroying the // docshell. If we are destroying it, then starting layout will // likely cause us to crash, or at best waste a lot of time as we // are just going to tear it down anyway. bool destroying = true; if (mDocShell) { mDocShell->IsBeingDestroyed(&destroying); } if (!destroying) { nsContentSink::StartLayout(false); } } ScrollToRef(); mDocument->RemoveObserver(this); if (!mParser) { // DidBuildModelImpl may cause mParser to be nulled out // Return early to avoid unblocking the onload event too many times. return NS_OK; } // We may not have called BeginLoad() if loading is terminated before // OnStartRequest call. if (mStarted) { mDocument->EndLoad(); } DropParserAndPerfHint(); #ifdef GATHER_DOCWRITE_STATISTICS printf("UNSAFE SCRIPTS: %d\n", sUnsafeDocWrites); printf("TOKENIZER-SAFE SCRIPTS: %d\n", sTokenSafeDocWrites); printf("TREEBUILDER-SAFE SCRIPTS: %d\n", sTreeSafeDocWrites); #endif #ifdef DEBUG_NS_HTML5_TREE_OP_EXECUTOR_FLUSH printf("MAX NOTIFICATION BATCH LEN: %d\n", sAppendBatchMaxSize); if (sAppendBatchExaminations != 0) { printf("AVERAGE SLOTS EXAMINED: %d\n", sAppendBatchSlotsExamined / sAppendBatchExaminations); } #endif return NS_OK; } NS_IMETHODIMP nsHtml5TreeOpExecutor::WillInterrupt() { NS_NOTREACHED("Don't call. For interface compat only."); return NS_ERROR_NOT_IMPLEMENTED; } NS_IMETHODIMP nsHtml5TreeOpExecutor::WillResume() { NS_NOTREACHED("Don't call. For interface compat only."); return NS_ERROR_NOT_IMPLEMENTED; } NS_IMETHODIMP nsHtml5TreeOpExecutor::SetParser(nsParserBase* aParser) { mParser = aParser; return NS_OK; } void nsHtml5TreeOpExecutor::FlushPendingNotifications(mozFlushType aType) { if (aType >= Flush_InterruptibleLayout) { // Bug 577508 / 253951 nsContentSink::StartLayout(true); } } nsISupports* nsHtml5TreeOpExecutor::GetTarget() { return mDocument; } nsresult nsHtml5TreeOpExecutor::MarkAsBroken(nsresult aReason) { NS_ASSERTION(NS_IsMainThread(), "Wrong thread!"); mBroken = aReason; if (mStreamParser) { mStreamParser->Terminate(); } // We are under memory pressure, but let's hope the following allocation // works out so that we get to terminate and clean up the parser from // a safer point. if (mParser) { // can mParser ever be null here? MOZ_ALWAYS_SUCCEEDS( NS_DispatchToMainThread(NewRunnableMethod(GetParser(), &nsHtml5Parser::Terminate))); } return aReason; } void FlushTimerCallback(nsITimer* aTimer, void* aClosure) { RefPtr ex = gBackgroundFlushList->popFirst(); if (ex) { ex->RunFlushLoop(); } if (gBackgroundFlushList && gBackgroundFlushList->isEmpty()) { delete gBackgroundFlushList; gBackgroundFlushList = nullptr; gFlushTimer->Cancel(); NS_RELEASE(gFlushTimer); } } void nsHtml5TreeOpExecutor::ContinueInterruptedParsingAsync() { if (!mDocument || !mDocument->IsInBackgroundWindow()) { nsCOMPtr flusher = new nsHtml5ExecutorReflusher(this); if (NS_FAILED(NS_DispatchToMainThread(flusher))) { NS_WARNING("failed to dispatch executor flush event"); } } else { if (!gBackgroundFlushList) { gBackgroundFlushList = new mozilla::LinkedList(); } if (!isInList()) { gBackgroundFlushList->insertBack(this); } if (!gFlushTimer) { nsCOMPtr t = do_CreateInstance("@mozilla.org/timer;1"); t.swap(gFlushTimer); // The timer value 50 should not hopefully slow down background pages too // much, yet lets event loop to process enough between ticks. // See bug 734015. gFlushTimer->InitWithNamedFuncCallback(FlushTimerCallback, nullptr, 50, nsITimer::TYPE_REPEATING_SLACK, "FlushTimerCallback"); } } } void nsHtml5TreeOpExecutor::FlushSpeculativeLoads() { nsTArray speculativeLoadQueue; mStage.MoveSpeculativeLoadsTo(speculativeLoadQueue); const nsHtml5SpeculativeLoad* start = speculativeLoadQueue.Elements(); const nsHtml5SpeculativeLoad* end = start + speculativeLoadQueue.Length(); for (nsHtml5SpeculativeLoad* iter = const_cast(start); iter < end; ++iter) { if (MOZ_UNLIKELY(!mParser)) { // An extension terminated the parser from a HTTP observer. return; } iter->Perform(this); } } class nsHtml5FlushLoopGuard { private: RefPtr mExecutor; #ifdef DEBUG_NS_HTML5_TREE_OP_EXECUTOR_FLUSH uint32_t mStartTime; #endif public: explicit nsHtml5FlushLoopGuard(nsHtml5TreeOpExecutor* aExecutor) : mExecutor(aExecutor) #ifdef DEBUG_NS_HTML5_TREE_OP_EXECUTOR_FLUSH , mStartTime(PR_IntervalToMilliseconds(PR_IntervalNow())) #endif { mExecutor->mRunFlushLoopOnStack = true; } ~nsHtml5FlushLoopGuard() { #ifdef DEBUG_NS_HTML5_TREE_OP_EXECUTOR_FLUSH uint32_t timeOffTheEventLoop = PR_IntervalToMilliseconds(PR_IntervalNow()) - mStartTime; if (timeOffTheEventLoop > nsHtml5TreeOpExecutor::sLongestTimeOffTheEventLoop) { nsHtml5TreeOpExecutor::sLongestTimeOffTheEventLoop = timeOffTheEventLoop; } printf("Longest time off the event loop: %d\n", nsHtml5TreeOpExecutor::sLongestTimeOffTheEventLoop); #endif mExecutor->mRunFlushLoopOnStack = false; } }; /** * The purpose of the loop here is to avoid returning to the main event loop */ void nsHtml5TreeOpExecutor::RunFlushLoop() { PROFILER_LABEL("nsHtml5TreeOpExecutor", "RunFlushLoop", js::ProfileEntry::Category::OTHER); if (mRunFlushLoopOnStack) { // There's already a RunFlushLoop() on the call stack. return; } nsHtml5FlushLoopGuard guard(this); // this is also the self-kungfu! RefPtr parserKungFuDeathGrip(mParser); RefPtr streamParserGrip; if (mParser) { streamParserGrip = GetParser()->GetStreamParser(); } mozilla::Unused << streamParserGrip; // Intentionally not used within function // Remember the entry time (void) nsContentSink::WillParseImpl(); for (;;) { if (!mParser) { // Parse has terminated. mOpQueue.Clear(); // clear in order to be able to assert in destructor return; } if (NS_FAILED(IsBroken())) { return; } if (!parserKungFuDeathGrip->IsParserEnabled()) { // The parser is blocked. return; } if (mFlushState != eNotFlushing) { // XXX Can this happen? In case it can, let's avoid crashing. return; } // If there are scripts executing, then the content sink is jumping the gun // (probably due to a synchronous XMLHttpRequest) and will re-enable us // later, see bug 460706. if (IsScriptExecuting()) { return; } if (mReadingFromStage) { nsTArray speculativeLoadQueue; mStage.MoveOpsAndSpeculativeLoadsTo(mOpQueue, speculativeLoadQueue); // Make sure speculative loads never start after the corresponding // normal loads for the same URLs. const nsHtml5SpeculativeLoad* start = speculativeLoadQueue.Elements(); const nsHtml5SpeculativeLoad* end = start + speculativeLoadQueue.Length(); for (nsHtml5SpeculativeLoad* iter = (nsHtml5SpeculativeLoad*)start; iter < end; ++iter) { iter->Perform(this); if (MOZ_UNLIKELY(!mParser)) { // An extension terminated the parser from a HTTP observer. mOpQueue.Clear(); // clear in order to be able to assert in destructor return; } } } else { FlushSpeculativeLoads(); // Make sure speculative loads never start after // the corresponding normal loads for the same // URLs. if (MOZ_UNLIKELY(!mParser)) { // An extension terminated the parser from a HTTP observer. mOpQueue.Clear(); // clear in order to be able to assert in destructor return; } // Now parse content left in the document.write() buffer queue if any. // This may generate tree ops on its own or dequeue a speculation. nsresult rv = GetParser()->ParseUntilBlocked(); if (NS_FAILED(rv)) { MarkAsBroken(rv); return; } } if (mOpQueue.IsEmpty()) { // Avoid bothering the rest of the engine with a doc update if there's // nothing to do. return; } mFlushState = eInFlush; nsIContent* scriptElement = nullptr; BeginDocUpdate(); uint32_t numberOfOpsToFlush = mOpQueue.Length(); const nsHtml5TreeOperation* first = mOpQueue.Elements(); const nsHtml5TreeOperation* last = first + numberOfOpsToFlush - 1; for (nsHtml5TreeOperation* iter = const_cast(first);;) { if (MOZ_UNLIKELY(!mParser)) { // The previous tree op caused a call to nsIParser::Terminate(). break; } NS_ASSERTION(mFlushState == eInDocUpdate, "Tried to perform tree op outside update batch."); nsresult rv = iter->Perform(this, &scriptElement); if (NS_FAILED(rv)) { MarkAsBroken(rv); break; } // Be sure not to check the deadline if the last op was just performed. if (MOZ_UNLIKELY(iter == last)) { break; } else if (MOZ_UNLIKELY(nsContentSink::DidProcessATokenImpl() == NS_ERROR_HTMLPARSER_INTERRUPTED)) { mOpQueue.RemoveElementsAt(0, (iter - first) + 1); EndDocUpdate(); mFlushState = eNotFlushing; #ifdef DEBUG_NS_HTML5_TREE_OP_EXECUTOR_FLUSH printf("REFLUSH SCHEDULED (executing ops): %d\n", ++sTimesFlushLoopInterrupted); #endif nsHtml5TreeOpExecutor::ContinueInterruptedParsingAsync(); return; } ++iter; } mOpQueue.Clear(); EndDocUpdate(); mFlushState = eNotFlushing; if (MOZ_UNLIKELY(!mParser)) { // The parse ended already. return; } if (scriptElement) { // must be tail call when mFlushState is eNotFlushing RunScript(scriptElement); // Always check the clock in nsContentSink right after a script StopDeflecting(); if (nsContentSink::DidProcessATokenImpl() == NS_ERROR_HTMLPARSER_INTERRUPTED) { #ifdef DEBUG_NS_HTML5_TREE_OP_EXECUTOR_FLUSH printf("REFLUSH SCHEDULED (after script): %d\n", ++sTimesFlushLoopInterrupted); #endif nsHtml5TreeOpExecutor::ContinueInterruptedParsingAsync(); return; } } } } nsresult nsHtml5TreeOpExecutor::FlushDocumentWrite() { nsresult rv = IsBroken(); NS_ENSURE_SUCCESS(rv, rv); FlushSpeculativeLoads(); // Make sure speculative loads never start after the // corresponding normal loads for the same URLs. if (MOZ_UNLIKELY(!mParser)) { // The parse has ended. mOpQueue.Clear(); // clear in order to be able to assert in destructor return rv; } if (mFlushState != eNotFlushing) { // XXX Can this happen? In case it can, let's avoid crashing. return rv; } mFlushState = eInFlush; // avoid crashing near EOF RefPtr kungFuDeathGrip(this); RefPtr parserKungFuDeathGrip(mParser); mozilla::Unused << parserKungFuDeathGrip; // Intentionally not used within function RefPtr streamParserGrip; if (mParser) { streamParserGrip = GetParser()->GetStreamParser(); } mozilla::Unused << streamParserGrip; // Intentionally not used within function NS_ASSERTION(!mReadingFromStage, "Got doc write flush when reading from stage"); #ifdef DEBUG mStage.AssertEmpty(); #endif nsIContent* scriptElement = nullptr; BeginDocUpdate(); uint32_t numberOfOpsToFlush = mOpQueue.Length(); const nsHtml5TreeOperation* start = mOpQueue.Elements(); const nsHtml5TreeOperation* end = start + numberOfOpsToFlush; for (nsHtml5TreeOperation* iter = const_cast(start); iter < end; ++iter) { if (MOZ_UNLIKELY(!mParser)) { // The previous tree op caused a call to nsIParser::Terminate(). break; } NS_ASSERTION(mFlushState == eInDocUpdate, "Tried to perform tree op outside update batch."); rv = iter->Perform(this, &scriptElement); if (NS_FAILED(rv)) { MarkAsBroken(rv); break; } } mOpQueue.Clear(); EndDocUpdate(); mFlushState = eNotFlushing; if (MOZ_UNLIKELY(!mParser)) { // Ending the doc update caused a call to nsIParser::Terminate(). return rv; } if (scriptElement) { // must be tail call when mFlushState is eNotFlushing RunScript(scriptElement); } return rv; } // copied from HTML content sink bool nsHtml5TreeOpExecutor::IsScriptEnabled() { // Note that if we have no document or no docshell or no global or whatnot we // want to claim script _is_ enabled, so we don't parse the contents of //