[ResourceScheduler] Yield after starting several requests

content/browser/loader/resource_scheduler.cc

 // Copyright (c) 2012 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.
 
 #include "content/browser/loader/resource_scheduler.h"
 
 #include <stdint.h>
 #include <set>
 #include <string>
 #include <utility>
 #include <vector>
 
 #include "base/feature_list.h"
 #include "base/macros.h"
 #include "base/metrics/field_trial.h"
+#include "base/metrics/field_trial_params.h"
 #include "base/metrics/histogram_macros.h"
 #include "base/stl_util.h"
 #include "base/supports_user_data.h"
 #include "base/threading/thread_task_runner_handle.h"
 #include "content/common/resource_messages.h"
 #include "content/public/browser/resource_request_info.h"
 #include "content/public/browser/resource_throttle.h"
 #include "net/base/host_port_pair.h"
 #include "net/base/load_flags.h"
 #include "net/base/request_priority.h"
 #include "net/http/http_server_properties.h"
 #include "net/url_request/url_request.h"
 #include "net/url_request/url_request_context.h"
 #include "url/scheme_host_port.h"
 
 namespace content {
 
 namespace {
 
 // When kPrioritySupportedRequestsDelayable is enabled, requests for
 // H2/QUIC/SPDY resources can be delayed by the ResourceScheduler just as
 // HTTP/1.1 resources are. Disabling this appears to have negative performance
 // impact, see https://crbug.com/655585.
 const base::Feature kPrioritySupportedRequestsDelayable{
     "PrioritySupportedRequestsDelayable", base::FEATURE_DISABLED_BY_DEFAULT};
 
+// In the event that many resource requests are started quickly, this feature
+// will periodically yield (e.g., delaying starting of requests) by posting a
+// task and waiting for the task to run to resume. This allows other
+// operations that rely on the IO thread (e.g., already running network
+// requests) to make progress.
+const base::Feature kNetworkSchedulerYielding{
+    "NetworkSchedulerYielding", base::FEATURE_DISABLED_BY_DEFAULT};
+const char kMaxRequestsBeforeYieldingParam[] = "MaxRequestsBeforeYieldingParam";
+const int kMaxRequestsBeforeYieldingDefault = 5;
+
 enum StartMode {
   START_SYNC,
   START_ASYNC
 };
 
 // Flags identifying various attributes of the request that are used
 // when making scheduling decisions.
 using RequestAttributes = uint8_t;
 const RequestAttributes kAttributeNone = 0x00;
 const RequestAttributes kAttributeInFlight = 0x01;
 const RequestAttributes kAttributeDelayable = 0x02;
 const RequestAttributes kAttributeLayoutBlocking = 0x04;
 
 // Reasons why pending requests may be started.  For logging only.
 enum class RequestStartTrigger {
   NONE,
   COMPLETION_PRE_BODY,
   COMPLETION_POST_BODY,
   BODY_REACHED,
   CLIENT_KILL,
   SPDY_PROXY_DETECTED,
   REQUEST_REPRIORITIZED,
+  START_YIELDED,

[Charlie Harrison, 2017/02/09 19:08:04]

nit: I think START_WAS_YIELDED is a bit better.

[jkarlin, 2017/02/09 19:48:12]

Done.

 };
 
 const char* RequestStartTriggerString(RequestStartTrigger trigger) {
   switch (trigger) {
     case RequestStartTrigger::NONE:
       return "NONE";
     case RequestStartTrigger::COMPLETION_PRE_BODY:
       return "COMPLETION_PRE_BODY";
     case RequestStartTrigger::COMPLETION_POST_BODY:
       return "COMPLETION_POST_BODY";
     case RequestStartTrigger::BODY_REACHED:
       return "BODY_REACHED";
     case RequestStartTrigger::CLIENT_KILL:
       return "CLIENT_KILL";
     case RequestStartTrigger::SPDY_PROXY_DETECTED:
       return "SPDY_PROXY_DETECTED";
     case RequestStartTrigger::REQUEST_REPRIORITIZED:
       return "REQUEST_REPRIORITIZED";
+    case RequestStartTrigger::START_YIELDED:
+      return "START_YIELDED";
   }
   NOTREACHED();
   return "Unknown";
 }
 
 }  // namespace
 
 // The maximum number of delayable requests to allow to be in-flight at any
 // point in time (across all hosts).
 static const size_t kMaxNumDelayableRequestsPerClient = 10;
@@ skipping 249 matching lines @@
 void ResourceScheduler::RequestQueue::Insert(
     ScheduledResourceRequest* request) {
   DCHECK(!base::ContainsKey(pointers_, request));
   request->set_fifo_ordering(MakeFifoOrderingId());
   pointers_[request] = queue_.insert(request);
 }
 
 // Each client represents a tab.
 class ResourceScheduler::Client {
  public:
-  explicit Client(bool priority_requests_delayable)
+  Client(bool priority_requests_delayable,
+         bool yielding_scheduler,
+         int max_requests_before_yielding)
       : is_loaded_(false),
         has_html_body_(false),
         using_spdy_proxy_(false),
         in_flight_delayable_count_(0),
         total_layout_blocking_count_(0),
         priority_requests_delayable_(priority_requests_delayable),
         has_pending_start_task_(false),
+        requests_since_yielding_(0),
+        yielding_scheduler_(yielding_scheduler),
+        max_requests_before_yielding_(max_requests_before_yielding),
         weak_ptr_factory_(this) {}
 
   ~Client() {}
 
   void ScheduleRequest(net::URLRequest* url_request,
                        ScheduledResourceRequest* request) {
     SetRequestAttributes(request, DetermineRequestAttributes(request));
-    if (ShouldStartRequest(request) == START_REQUEST) {
+    if (ShouldStartRequest(request) == START_REQUEST &&

[Charlie Harrison, 2017/02/09 19:08:04]

Eek, I am a little worried about sync XHRs here. Maybe not a big deal in
practice?

[jkarlin, 2017/02/09 19:48:13]

Nice catch.  I've changed it to not yield for sync requests and added a test.

+        !ShouldYieldScheduler()) {
       // New requests can be started synchronously without issue.
       StartRequest(request, START_SYNC, RequestStartTrigger::NONE);
     } else {
       pending_requests_.Insert(request);
     }
   }
 
   void RemoveRequest(ScheduledResourceRequest* request) {
     if (pending_requests_.IsQueued(request)) {
       pending_requests_.Erase(request);
@@ skipping 354 matching lines @@
   // pattern.
   void ScheduleLoadAnyStartablePendingRequests(RequestStartTrigger trigger) {
     if (has_pending_start_task_)
       return;
     has_pending_start_task_ = true;
     base::ThreadTaskRunnerHandle::Get()->PostTask(
         FROM_HERE, base::Bind(&Client::LoadAnyStartablePendingRequests,
                               weak_ptr_factory_.GetWeakPtr(), trigger));
   }
 
+  void ResumeAfterYielding() {
+    requests_since_yielding_ = 0;

[Charlie Harrison, 2017/02/09 19:08:04]

// A single request can trigger this method, so don't continue searching 
// unless we really yielded.
bool should_search_for_requests = requests_since_yielding_ >
max_requests_before_yielding_;
requests_since_yielding_ = 0;
if (should_search_for_requests)
  LoadAnyStartable...

[jkarlin, 2017/02/09 19:48:13]

Thanks, I forgot to add that back when I switched from starting task when first
yielding to first request.

+    LoadAnyStartablePendingRequests(RequestStartTrigger::START_YIELDED);
+  }
+
+  // Returns true if the scheduler should not start the next request and
+  // instead return so that other tasks can run on the IO thread (e.g.,
+  // existing network requests). If it returns true, ShouldYieldScheduler will
+  // have already scheduled a task to resume after yielding.
+  bool ShouldYieldScheduler() {
+    if (!yielding_scheduler_)
+      return false;
+
+    if (requests_since_yielding_ == 0) {

[Charlie Harrison, 2017/02/09 19:08:04]

The fact that we post a task immediately deserves some explanation.

[jkarlin, 2017/02/09 19:48:12]

Done.

+      base::ThreadTaskRunnerHandle::Get()->PostTask(
+          FROM_HERE, base::Bind(&Client::ResumeAfterYielding,
+                                weak_ptr_factory_.GetWeakPtr()));
+    }
+
+    return requests_since_yielding_++ >= max_requests_before_yielding_;

[Charlie Harrison, 2017/02/09 19:08:04]

optional: Can you put the increment on a separate line? This adds cognitive
burden for me.

[jkarlin, 2017/02/09 19:48:12]

Done.

+  }
+
   void LoadAnyStartablePendingRequests(RequestStartTrigger trigger) {
     // We iterate through all the pending requests, starting with the highest
     // priority one. For each entry, one of three things can happen:
     // 1) We start the request, remove it from the list, and keep checking.
     // 2) We do NOT start the request, but ShouldStartRequest() signals us that
     //     there may be room for other requests, so we keep checking and leave
     //     the previous request still in the list.
     // 3) We do not start the request, same as above, but StartRequest() tells
     //     us there's no point in checking any further requests.
     has_pending_start_task_ = false;
     RequestQueue::NetQueue::iterator request_iter =
         pending_requests_.GetNextHighestIterator();
 
     while (request_iter != pending_requests_.End()) {
       ScheduledResourceRequest* request = *request_iter;
       ShouldStartReqResult query_result = ShouldStartRequest(request);
 
       if (query_result == START_REQUEST) {
+        if (ShouldYieldScheduler())
+          break;
+
         pending_requests_.Erase(request);
         StartRequest(request, START_ASYNC, trigger);
 
         // StartRequest can modify the pending list, so we (re)start evaluation
         // from the currently highest priority request. Avoid copying a singular
         // iterator, which would trigger undefined behavior.
         if (pending_requests_.GetNextHighestIterator() ==
             pending_requests_.End())
           break;
         request_iter = pending_requests_.GetNextHighestIterator();
@@ skipping 18 matching lines @@
   size_t in_flight_delayable_count_;
   // The number of layout-blocking in-flight requests.
   size_t total_layout_blocking_count_;
 
   // True if requests to servers that support priorities (e.g., H2/QUIC) can
   // be delayed.
   bool priority_requests_delayable_;
 
   bool has_pending_start_task_;
 
+  // The number of requests that have been started since the last
+  // ResumeAfterYielding task was posted.
+  int requests_since_yielding_;
+
+  // Whether or not to periodically yield when starting lots of requests.
+  bool yielding_scheduler_;
+
+  // The number of requests that can start before yielding.
+  int max_requests_before_yielding_;
+
   base::WeakPtrFactory<ResourceScheduler::Client> weak_ptr_factory_;
 };
 
 ResourceScheduler::ResourceScheduler()
     : priority_requests_delayable_(
-          base::FeatureList::IsEnabled(kPrioritySupportedRequestsDelayable)) {}
+          base::FeatureList::IsEnabled(kPrioritySupportedRequestsDelayable)),
+      yielding_scheduler_(
+          base::FeatureList::IsEnabled(kNetworkSchedulerYielding)),
+      max_requests_before_yielding_(base::GetFieldTrialParamByFeatureAsInt(
+          kNetworkSchedulerYielding,
+          kMaxRequestsBeforeYieldingParam,
+          kMaxRequestsBeforeYieldingDefault)) {}
 
 ResourceScheduler::~ResourceScheduler() {
   DCHECK(unowned_requests_.empty());
   DCHECK(client_map_.empty());
 }
 
 std::unique_ptr<ResourceThrottle> ResourceScheduler::ScheduleRequest(
     int child_id,
     int route_id,
     bool is_async,
@@ skipping 36 matching lines @@
   Client* client = client_it->second;
   client->RemoveRequest(request);
 }
 
 void ResourceScheduler::OnClientCreated(int child_id,
                                         int route_id) {
   DCHECK(CalledOnValidThread());
   ClientId client_id = MakeClientId(child_id, route_id);
   DCHECK(!base::ContainsKey(client_map_, client_id));
 
-  Client* client = new Client(priority_requests_delayable_);
+  Client* client = new Client(priority_requests_delayable_, yielding_scheduler_,
+                              max_requests_before_yielding_);
   client_map_[client_id] = client;
 }
 
 void ResourceScheduler::OnClientDeleted(int child_id, int route_id) {
   DCHECK(CalledOnValidThread());
   ClientId client_id = MakeClientId(child_id, route_id);
   ClientMap::iterator it = client_map_.find(client_id);
   DCHECK(it != client_map_.end());
 
   Client* client = it->second;
@@ skipping 118 matching lines @@
   client->ReprioritizeRequest(scheduled_resource_request, old_priority_params,
                               new_priority_params);
 }
 
 ResourceScheduler::ClientId ResourceScheduler::MakeClientId(
     int child_id, int route_id) {
   return (static_cast<ResourceScheduler::ClientId>(child_id) << 32) | route_id;
 }
 
 }  // namespace content