1 // Copyright (c) 2012 The Chromium Authors. All rights reserved.
2 // Use of this source code is governed by a BSD-style license that can be
3 // found in the LICENSE file.
5 #include "net/dns/dns_session.h"
7 #include "base/basictypes.h"
9 #include "base/lazy_instance.h"
10 #include "base/metrics/histogram.h"
11 #include "base/metrics/sample_vector.h"
12 #include "base/rand_util.h"
13 #include "base/stl_util.h"
14 #include "base/time/time.h"
15 #include "net/base/ip_endpoint.h"
16 #include "net/base/net_errors.h"
17 #include "net/dns/dns_config_service.h"
18 #include "net/dns/dns_socket_pool.h"
19 #include "net/socket/stream_socket.h"
20 #include "net/udp/datagram_client_socket.h"
25 // Never exceed max timeout.
26 const unsigned kMaxTimeoutMs
= 5000;
27 // Set min timeout, in case we are talking to a local DNS proxy.
28 const unsigned kMinTimeoutMs
= 10;
30 // Number of buckets in the histogram of observed RTTs.
31 const size_t kRTTBucketCount
= 100;
32 // Target percentile in the RTT histogram used for retransmission timeout.
33 const unsigned kRTOPercentile
= 99;
36 // Runtime statistics of DNS server.
37 struct DnsSession::ServerStats
{
38 ServerStats(base::TimeDelta rtt_estimate_param
, RttBuckets
* buckets
)
39 : last_failure_count(0), rtt_estimate(rtt_estimate_param
) {
40 rtt_histogram
.reset(new base::SampleVector(buckets
));
41 // Seed histogram with 2 samples at |rtt_estimate| timeout.
42 rtt_histogram
->Accumulate(
43 static_cast<base::HistogramBase::Sample
>(rtt_estimate
.InMilliseconds()),
47 // Count of consecutive failures after last success.
48 int last_failure_count
;
50 // Last time when server returned failure or timeout.
51 base::Time last_failure
;
52 // Last time when server returned success.
53 base::Time last_success
;
55 // Estimated RTT using moving average.
56 base::TimeDelta rtt_estimate
;
57 // Estimated error in the above.
58 base::TimeDelta rtt_deviation
;
60 // A histogram of observed RTT .
61 scoped_ptr
<base::SampleVector
> rtt_histogram
;
63 DISALLOW_COPY_AND_ASSIGN(ServerStats
);
67 base::LazyInstance
<DnsSession::RttBuckets
>::Leaky
DnsSession::rtt_buckets_
=
68 LAZY_INSTANCE_INITIALIZER
;
70 DnsSession::RttBuckets::RttBuckets() : base::BucketRanges(kRTTBucketCount
+ 1) {
71 base::Histogram::InitializeBucketRanges(1, 5000, this);
74 DnsSession::SocketLease::SocketLease(scoped_refptr
<DnsSession
> session
,
75 unsigned server_index
,
76 scoped_ptr
<DatagramClientSocket
> socket
)
77 : session_(session
), server_index_(server_index
), socket_(socket
.Pass()) {}
79 DnsSession::SocketLease::~SocketLease() {
80 session_
->FreeSocket(server_index_
, socket_
.Pass());
83 DnsSession::DnsSession(const DnsConfig
& config
,
84 scoped_ptr
<DnsSocketPool
> socket_pool
,
85 const RandIntCallback
& rand_int_callback
,
88 socket_pool_(socket_pool
.Pass()),
89 rand_callback_(base::Bind(rand_int_callback
, 0, kuint16max
)),
92 socket_pool_
->Initialize(&config_
.nameservers
, net_log
);
93 UMA_HISTOGRAM_CUSTOM_COUNTS(
94 "AsyncDNS.ServerCount", config_
.nameservers
.size(), 0, 10, 11);
95 for (size_t i
= 0; i
< config_
.nameservers
.size(); ++i
) {
96 server_stats_
.push_back(new ServerStats(config_
.timeout
,
97 rtt_buckets_
.Pointer()));
101 DnsSession::~DnsSession() {
105 uint16
DnsSession::NextQueryId() const {
106 return static_cast<uint16
>(rand_callback_
.Run());
109 unsigned DnsSession::NextFirstServerIndex() {
110 unsigned index
= NextGoodServerIndex(server_index_
);
112 server_index_
= (server_index_
+ 1) % config_
.nameservers
.size();
116 unsigned DnsSession::NextGoodServerIndex(unsigned server_index
) {
117 unsigned index
= server_index
;
118 base::Time
oldest_server_failure(base::Time::Now());
119 unsigned oldest_server_failure_index
= 0;
121 UMA_HISTOGRAM_BOOLEAN("AsyncDNS.ServerIsGood",
122 server_stats_
[server_index
]->last_failure
.is_null());
125 base::Time cur_server_failure
= server_stats_
[index
]->last_failure
;
126 // If number of failures on this server doesn't exceed number of allowed
127 // attempts, return its index.
128 if (server_stats_
[server_index
]->last_failure_count
< config_
.attempts
) {
131 // Track oldest failed server.
132 if (cur_server_failure
< oldest_server_failure
) {
133 oldest_server_failure
= cur_server_failure
;
134 oldest_server_failure_index
= index
;
136 index
= (index
+ 1) % config_
.nameservers
.size();
137 } while (index
!= server_index
);
139 // If we are here it means that there are no successful servers, so we have
140 // to use one that has failed oldest.
141 return oldest_server_failure_index
;
144 void DnsSession::RecordServerFailure(unsigned server_index
) {
145 UMA_HISTOGRAM_CUSTOM_COUNTS(
146 "AsyncDNS.ServerFailureIndex", server_index
, 0, 10, 11);
147 ++(server_stats_
[server_index
]->last_failure_count
);
148 server_stats_
[server_index
]->last_failure
= base::Time::Now();
151 void DnsSession::RecordServerSuccess(unsigned server_index
) {
152 if (server_stats_
[server_index
]->last_success
.is_null()) {
153 UMA_HISTOGRAM_COUNTS_100("AsyncDNS.ServerFailuresAfterNetworkChange",
154 server_stats_
[server_index
]->last_failure_count
);
156 UMA_HISTOGRAM_COUNTS_100("AsyncDNS.ServerFailuresBeforeSuccess",
157 server_stats_
[server_index
]->last_failure_count
);
159 server_stats_
[server_index
]->last_failure_count
= 0;
160 server_stats_
[server_index
]->last_failure
= base::Time();
161 server_stats_
[server_index
]->last_success
= base::Time::Now();
164 void DnsSession::RecordRTT(unsigned server_index
, base::TimeDelta rtt
) {
165 DCHECK_LT(server_index
, server_stats_
.size());
167 // For measurement, assume it is the first attempt (no backoff).
168 base::TimeDelta timeout_jacobson
= NextTimeoutFromJacobson(server_index
, 0);
169 base::TimeDelta timeout_histogram
= NextTimeoutFromHistogram(server_index
, 0);
170 UMA_HISTOGRAM_TIMES("AsyncDNS.TimeoutErrorJacobson", rtt
- timeout_jacobson
);
171 UMA_HISTOGRAM_TIMES("AsyncDNS.TimeoutErrorHistogram",
172 rtt
- timeout_histogram
);
173 UMA_HISTOGRAM_TIMES("AsyncDNS.TimeoutErrorJacobsonUnder",
174 timeout_jacobson
- rtt
);
175 UMA_HISTOGRAM_TIMES("AsyncDNS.TimeoutErrorHistogramUnder",
176 timeout_histogram
- rtt
);
178 // Jacobson/Karels algorithm for TCP.
179 // Using parameters: alpha = 1/8, delta = 1/4, beta = 4
180 base::TimeDelta
& estimate
= server_stats_
[server_index
]->rtt_estimate
;
181 base::TimeDelta
& deviation
= server_stats_
[server_index
]->rtt_deviation
;
182 base::TimeDelta current_error
= rtt
- estimate
;
183 estimate
+= current_error
/ 8; // * alpha
184 base::TimeDelta abs_error
= base::TimeDelta::FromInternalValue(
185 std::abs(current_error
.ToInternalValue()));
186 deviation
+= (abs_error
- deviation
) / 4; // * delta
188 // Histogram-based method.
189 server_stats_
[server_index
]->rtt_histogram
->Accumulate(
190 static_cast<base::HistogramBase::Sample
>(rtt
.InMilliseconds()), 1);
193 void DnsSession::RecordLostPacket(unsigned server_index
, int attempt
) {
194 base::TimeDelta timeout_jacobson
=
195 NextTimeoutFromJacobson(server_index
, attempt
);
196 base::TimeDelta timeout_histogram
=
197 NextTimeoutFromHistogram(server_index
, attempt
);
198 UMA_HISTOGRAM_TIMES("AsyncDNS.TimeoutSpentJacobson", timeout_jacobson
);
199 UMA_HISTOGRAM_TIMES("AsyncDNS.TimeoutSpentHistogram", timeout_histogram
);
202 void DnsSession::RecordServerStats() {
203 for (size_t index
= 0; index
< server_stats_
.size(); ++index
) {
204 if (server_stats_
[index
]->last_failure_count
) {
205 if (server_stats_
[index
]->last_success
.is_null()) {
206 UMA_HISTOGRAM_COUNTS("AsyncDNS.ServerFailuresWithoutSuccess",
207 server_stats_
[index
]->last_failure_count
);
209 UMA_HISTOGRAM_COUNTS("AsyncDNS.ServerFailuresAfterSuccess",
210 server_stats_
[index
]->last_failure_count
);
217 base::TimeDelta
DnsSession::NextTimeout(unsigned server_index
, int attempt
) {
218 // Respect config timeout if it exceeds |kMaxTimeoutMs|.
219 if (config_
.timeout
.InMilliseconds() >= kMaxTimeoutMs
)
220 return config_
.timeout
;
221 return NextTimeoutFromHistogram(server_index
, attempt
);
224 // Allocate a socket, already connected to the server address.
225 scoped_ptr
<DnsSession::SocketLease
> DnsSession::AllocateSocket(
226 unsigned server_index
, const NetLog::Source
& source
) {
227 scoped_ptr
<DatagramClientSocket
> socket
;
229 socket
= socket_pool_
->AllocateSocket(server_index
);
231 return scoped_ptr
<SocketLease
>();
233 socket
->NetLog().BeginEvent(NetLog::TYPE_SOCKET_IN_USE
,
234 source
.ToEventParametersCallback());
236 SocketLease
* lease
= new SocketLease(this, server_index
, socket
.Pass());
237 return scoped_ptr
<SocketLease
>(lease
);
240 scoped_ptr
<StreamSocket
> DnsSession::CreateTCPSocket(
241 unsigned server_index
, const NetLog::Source
& source
) {
242 return socket_pool_
->CreateTCPSocket(server_index
, source
);
246 void DnsSession::FreeSocket(unsigned server_index
,
247 scoped_ptr
<DatagramClientSocket
> socket
) {
248 DCHECK(socket
.get());
250 socket
->NetLog().EndEvent(NetLog::TYPE_SOCKET_IN_USE
);
252 socket_pool_
->FreeSocket(server_index
, socket
.Pass());
255 base::TimeDelta
DnsSession::NextTimeoutFromJacobson(unsigned server_index
,
257 DCHECK_LT(server_index
, server_stats_
.size());
259 base::TimeDelta timeout
= server_stats_
[server_index
]->rtt_estimate
+
260 4 * server_stats_
[server_index
]->rtt_deviation
;
262 timeout
= std::max(timeout
, base::TimeDelta::FromMilliseconds(kMinTimeoutMs
));
264 // The timeout doubles every full round.
265 unsigned num_backoffs
= attempt
/ config_
.nameservers
.size();
267 return std::min(timeout
* (1 << num_backoffs
),
268 base::TimeDelta::FromMilliseconds(kMaxTimeoutMs
));
271 base::TimeDelta
DnsSession::NextTimeoutFromHistogram(unsigned server_index
,
273 DCHECK_LT(server_index
, server_stats_
.size());
275 static_assert(std::numeric_limits
<base::HistogramBase::Count
>::is_signed
,
276 "histogram base count assumed to be signed");
278 // Use fixed percentile of observed samples.
279 const base::SampleVector
& samples
=
280 *server_stats_
[server_index
]->rtt_histogram
;
282 base::HistogramBase::Count total
= samples
.TotalCount();
283 base::HistogramBase::Count remaining_count
= kRTOPercentile
* total
/ 100;
285 while (remaining_count
> 0 && index
< rtt_buckets_
.Get().size()) {
286 remaining_count
-= samples
.GetCountAtIndex(index
);
290 base::TimeDelta timeout
=
291 base::TimeDelta::FromMilliseconds(rtt_buckets_
.Get().range(index
));
293 timeout
= std::max(timeout
, base::TimeDelta::FromMilliseconds(kMinTimeoutMs
));
295 // The timeout still doubles every full round.
296 unsigned num_backoffs
= attempt
/ config_
.nameservers
.size();
298 return std::min(timeout
* (1 << num_backoffs
),
299 base::TimeDelta::FromMilliseconds(kMaxTimeoutMs
));