OLD | NEW |
(Empty) | |
| 1 // Copyright 2016 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. |
| 4 |
| 5 #include "net/filter/sdch_policy_delegate.h" |
| 6 |
| 7 #include "base/metrics/histogram_macros.h" |
| 8 #include "base/strings/string_util.h" |
| 9 #include "base/values.h" |
| 10 #include "net/base/sdch_problem_codes.h" |
| 11 #include "net/log/net_log.h" |
| 12 #include "net/log/net_log_event_type.h" |
| 13 #include "net/url_request/url_request_http_job.h" |
| 14 |
| 15 namespace net { |
| 16 |
| 17 namespace { |
| 18 |
| 19 const char kRefreshHtml[] = |
| 20 "<head><META HTTP-EQUIV=\"Refresh\" CONTENT=\"0\"></head>"; |
| 21 // Mime types: |
| 22 const char kTextHtml[] = "text/html"; |
| 23 |
| 24 // Disambiguate various types of responses that trigger a meta-refresh, |
| 25 // failure, or fallback to pass-through. |
| 26 enum ResponseCorruptionDetectionCause { |
| 27 RESPONSE_NONE, |
| 28 |
| 29 // 404 Http Response Code |
| 30 RESPONSE_404 = 1, |
| 31 |
| 32 // Not a 200 Http Response Code |
| 33 RESPONSE_NOT_200 = 2, |
| 34 |
| 35 // Cached before dictionary retrieved. |
| 36 RESPONSE_OLD_UNENCODED = 3, |
| 37 |
| 38 // Speculative but incorrect SDCH filtering was added added. |
| 39 RESPONSE_TENTATIVE_SDCH = 4, |
| 40 |
| 41 // Missing correct dict for decoding. |
| 42 RESPONSE_NO_DICTIONARY = 5, |
| 43 |
| 44 // Not an SDCH response but should be. |
| 45 RESPONSE_CORRUPT_SDCH = 6, |
| 46 |
| 47 // No dictionary was advertised with the request, the server claims |
| 48 // to have encoded with SDCH anyway, but it isn't an SDCH response. |
| 49 RESPONSE_ENCODING_LIE = 7, |
| 50 |
| 51 RESPONSE_MAX, |
| 52 }; |
| 53 |
| 54 const char* ResponseCorruptionDetectionCauseToString( |
| 55 ResponseCorruptionDetectionCause cause) { |
| 56 const char* cause_string = "<unknown>"; |
| 57 switch (cause) { |
| 58 case RESPONSE_NONE: |
| 59 cause_string = "NONE"; |
| 60 break; |
| 61 case RESPONSE_404: |
| 62 cause_string = "404"; |
| 63 break; |
| 64 case RESPONSE_NOT_200: |
| 65 cause_string = "NOT_200"; |
| 66 break; |
| 67 case RESPONSE_OLD_UNENCODED: |
| 68 cause_string = "OLD_UNENCODED"; |
| 69 break; |
| 70 case RESPONSE_TENTATIVE_SDCH: |
| 71 cause_string = "TENTATIVE_SDCH"; |
| 72 break; |
| 73 case RESPONSE_NO_DICTIONARY: |
| 74 cause_string = "NO_DICTIONARY"; |
| 75 break; |
| 76 case RESPONSE_CORRUPT_SDCH: |
| 77 cause_string = "CORRUPT_SDCH"; |
| 78 break; |
| 79 case RESPONSE_ENCODING_LIE: |
| 80 cause_string = "ENCODING_LIE"; |
| 81 break; |
| 82 case RESPONSE_MAX: |
| 83 cause_string = "<Error: max enum value>"; |
| 84 break; |
| 85 } |
| 86 return cause_string; |
| 87 } |
| 88 |
| 89 std::unique_ptr<base::Value> NetLogResponseCorruptionDetectionCallback( |
| 90 ResponseCorruptionDetectionCause cause, |
| 91 bool cached, |
| 92 NetLogCaptureMode capture_mode) { |
| 93 std::unique_ptr<base::DictionaryValue> dict(new base::DictionaryValue()); |
| 94 dict->SetString("cause", ResponseCorruptionDetectionCauseToString(cause)); |
| 95 dict->SetBoolean("cached", cached); |
| 96 return std::move(dict); |
| 97 } |
| 98 |
| 99 void LogCorruptionDetection(const NetLogWithSource& net_log, |
| 100 bool is_cached_content, |
| 101 ResponseCorruptionDetectionCause cause) { |
| 102 // Use if statement rather than ?: because UMA_HISTOGRAM_ENUMERATION |
| 103 // caches the histogram name based on the call site. |
| 104 if (is_cached_content) { |
| 105 UMA_HISTOGRAM_ENUMERATION("Sdch3.ResponseCorruptionDetection.Cached", cause, |
| 106 RESPONSE_MAX); |
| 107 } else { |
| 108 UMA_HISTOGRAM_ENUMERATION("Sdch3.ResponseCorruptionDetection.Uncached", |
| 109 cause, RESPONSE_MAX); |
| 110 } |
| 111 net_log.AddEvent(NetLogEventType::SDCH_RESPONSE_CORRUPTION_DETECTION, |
| 112 base::Bind(&NetLogResponseCorruptionDetectionCallback, cause, |
| 113 is_cached_content)); |
| 114 } |
| 115 |
| 116 } // namespace |
| 117 |
| 118 SdchPolicyDelegate::SdchPolicyDelegate( |
| 119 bool possible_pass_through, |
| 120 URLRequestHttpJob* job, |
| 121 std::string mime_type, |
| 122 const GURL& url, |
| 123 bool is_cached_content, |
| 124 SdchManager* sdch_manager, |
| 125 std::unique_ptr<SdchManager::DictionarySet> dictionary_set, |
| 126 int response_code, |
| 127 const NetLogWithSource& net_log) |
| 128 : possible_pass_through_(possible_pass_through), |
| 129 job_(job), |
| 130 mime_type_(mime_type), |
| 131 url_(url), |
| 132 is_cached_content_(is_cached_content), |
| 133 sdch_manager_(sdch_manager), |
| 134 dictionary_set_(std::move(dictionary_set)), |
| 135 response_code_(response_code), |
| 136 net_log_(net_log) {} |
| 137 |
| 138 SdchPolicyDelegate::~SdchPolicyDelegate() {} |
| 139 |
| 140 // static |
| 141 void SdchPolicyDelegate::FixUpSdchContentEncodings( |
| 142 const NetLogWithSource& net_log, |
| 143 const std::string& mime_type, |
| 144 SdchManager::DictionarySet* dictionary_set, |
| 145 std::vector<SourceStream::SourceType>* types) { |
| 146 if (!dictionary_set) { |
| 147 // If sdch dictionary is advertised, we might need to add some decoding, as |
| 148 // some proxies strip encoding completely. |
| 149 // It was not an SDCH request, so we'll just record stats. |
| 150 if (1 < types->size()) { |
| 151 // Multiple filters were intended to only be used for SDCH (thus far!) |
| 152 SdchManager::LogSdchProblem(net_log, |
| 153 SDCH_MULTIENCODING_FOR_NON_SDCH_REQUEST); |
| 154 } |
| 155 if ((1 == types->size()) && (SourceStream::TYPE_SDCH == types->front())) { |
| 156 SdchManager::LogSdchProblem( |
| 157 net_log, SDCH_SDCH_CONTENT_ENCODE_FOR_NON_SDCH_REQUEST); |
| 158 } |
| 159 return; |
| 160 } |
| 161 // If content encoding included SDCH, then everything is "relatively" fine. |
| 162 if (!types->empty() && types->front() == SourceStream::TYPE_SDCH) { |
| 163 // Some proxies (found currently in Argentina) strip the Content-Encoding |
| 164 // text from "sdch,gzip" to a mere "sdch" without modifying the compressed |
| 165 // payload. To handle this gracefully, we simulate the "probably" deleted |
| 166 // ",gzip" by appending a tentative gzip decode, which will default to a |
| 167 // no-op pass through filter if it doesn't get gzip headers where |
| 168 // expected. |
| 169 if (1 == types->size()) { |
| 170 types->push_back(SourceStream::TYPE_GZIP_FALLBACK); |
| 171 SdchManager::LogSdchProblem(net_log, SDCH_OPTIONAL_GUNZIP_ENCODING_ADDED); |
| 172 } |
| 173 return; |
| 174 } |
| 175 |
| 176 // There are now several cases to handle for an SDCH request. Foremost, if |
| 177 // the outbound request was stripped so as not to advertise support for |
| 178 // encodings, we might get back content with no encoding, or (for example) |
| 179 // just gzip. We have to be sure that any changes we make allow for such |
| 180 // minimal coding to work. That issue is why we use POSSIBLE_SDCH filters, |
| 181 // as those filters sniff the content, and act as pass-through filters if |
| 182 // headers are not found. TODO(xunjieli): Currently POSSIBLE_SDCH filters |
| 183 // issue meta-refresh instead of pass-throughs to match old code. Change it |
| 184 // to do pass-throughs. |
| 185 |
| 186 // If the outbound GET is not modified, then the server will generally try to |
| 187 // send us SDCH encoded content. As that content returns, there are several |
| 188 // corruptions of the header "content-encoding" that proxies may perform (and |
| 189 // have been detected in the wild). We already dealt with the a honest |
| 190 // content encoding of "sdch,gzip" being corrupted into "sdch" with no change |
| 191 // of the actual content. Another common corruption is to either discard |
| 192 // the accurate content encoding, or to replace it with gzip only (again, with |
| 193 // no change in actual content). The last observed corruption it to actually |
| 194 // change the content, such as by re-gzipping it, and that may happen along |
| 195 // with corruption of the stated content encoding (wow!). |
| 196 |
| 197 // The one unresolved failure mode comes when we advertise a dictionary, and |
| 198 // the server tries to *send* a gzipped file (not gzip encode content), and |
| 199 // then we could do a gzip decode :-(. |
| 200 // We will gather a lot of stats as we perform the fixups. |
| 201 if (base::StartsWith(mime_type, kTextHtml, |
| 202 base::CompareCase::INSENSITIVE_ASCII)) { |
| 203 // Suspicious case: Advertised dictionary, but server didn't use sdch, and |
| 204 // we're HTML tagged. |
| 205 if (types->empty()) { |
| 206 SdchManager::LogSdchProblem(net_log, SDCH_ADDED_CONTENT_ENCODING); |
| 207 } else if (1 == types->size()) { |
| 208 SdchManager::LogSdchProblem(net_log, SDCH_FIXED_CONTENT_ENCODING); |
| 209 } else { |
| 210 SdchManager::LogSdchProblem(net_log, SDCH_FIXED_CONTENT_ENCODINGS); |
| 211 } |
| 212 } else { |
| 213 // Remarkable case!?! We advertised an SDCH dictionary, content-encoding |
| 214 // was not marked for SDCH processing: Why did the server suggest an SDCH |
| 215 // dictionary in the first place??. Also, the content isn't |
| 216 // tagged as HTML, despite the fact that SDCH encoding is mostly likely |
| 217 // for HTML: Did some anti-virus system strip this tag (sometimes they |
| 218 // strip accept-encoding headers on the request)?? Does the content |
| 219 // encoding not start with "text/html" for some other reason?? We'll |
| 220 // report this as a fixup to a binary file, but it probably really is |
| 221 // text/html (some how). |
| 222 if (types->empty()) { |
| 223 SdchManager::LogSdchProblem(net_log, SDCH_BINARY_ADDED_CONTENT_ENCODING); |
| 224 } else if (1 == types->size()) { |
| 225 SdchManager::LogSdchProblem(net_log, SDCH_BINARY_FIXED_CONTENT_ENCODING); |
| 226 } else { |
| 227 SdchManager::LogSdchProblem(net_log, SDCH_BINARY_FIXED_CONTENT_ENCODINGS); |
| 228 } |
| 229 } |
| 230 |
| 231 // Leave the existing encoding type to be processed first, and add our |
| 232 // tentative decodings to be done afterwards. Vodaphone UK reportedly will |
| 233 // perform a second layer of gzip encoding atop the server's sdch,gzip |
| 234 // encoding, and then claim that the content encoding is a mere gzip. As a |
| 235 // result we'll need (in that case) to do the gunzip, plus our tentative |
| 236 // gunzip and tentative SDCH decoding. This approach nicely handles the |
| 237 // empty() list as well, and should work with other (as yet undiscovered) |
| 238 // proxies the choose to re-compressed with some other encoding (such as |
| 239 // bzip2, etc.). |
| 240 types->insert(types->begin(), SourceStream::TYPE_GZIP_FALLBACK); |
| 241 types->insert(types->begin(), SourceStream::TYPE_SDCH_POSSIBLE); |
| 242 } |
| 243 |
| 244 // Dictionary id errors are often the first indication that the SDCH stream has |
| 245 // become corrupt. There are many possible causes: non-200 response codes, a |
| 246 // cached non-SDCH-ified reply, or a response that claims to be SDCH but isn't |
| 247 // actually. These are handled here by issuing a meta-refresh or swapping to the |
| 248 // "passthrough" mode if appropriate, or failing the request if the error is |
| 249 // unrecoverable. |
| 250 SdchPolicyDelegate::ErrorRecovery SdchPolicyDelegate::OnDictionaryIdError( |
| 251 std::string* replace_output) { |
| 252 if (possible_pass_through_) { |
| 253 LogCorruptionDetection(net_log_, is_cached_content_, |
| 254 RESPONSE_TENTATIVE_SDCH); |
| 255 // Ideally we should return PASS_THROUGH here, but this is done to match |
| 256 // the old behavior in sdch_filter.cc. |
| 257 } |
| 258 // HTTP 404 might be an unencoded error page, so if decoding failed, pass it |
| 259 // through. TODO(xunjieli): Remove this. crbug.com/516773. |
| 260 if (response_code_ == 404) { |
| 261 SdchManager::LogSdchProblem(net_log_, SDCH_PASS_THROUGH_404_CODE); |
| 262 LogCorruptionDetection(net_log_, is_cached_content_, RESPONSE_404); |
| 263 return PASS_THROUGH; |
| 264 } |
| 265 |
| 266 // HTTP !200 gets a meta-refresh for HTML. |
| 267 // TODO(xunjieli): remove this. crbug.com/654393. |
| 268 if (response_code_ != 200) { |
| 269 LogCorruptionDetection(net_log_, is_cached_content_, RESPONSE_NOT_200); |
| 270 return IssueMetaRefreshIfPossible(replace_output); |
| 271 } |
| 272 |
| 273 // If this is a cached result and the source hasn't requested a dictionary, it |
| 274 // probably never had a dictionary to begin and is an unencoded response from |
| 275 // earlier. |
| 276 if (is_cached_content_) { |
| 277 SdchManager::LogSdchProblem(net_log_, SDCH_PASS_THROUGH_OLD_CACHED); |
| 278 LogCorruptionDetection(net_log_, is_cached_content_, |
| 279 RESPONSE_OLD_UNENCODED); |
| 280 return PASS_THROUGH; |
| 281 } |
| 282 |
| 283 // The original request didn't advertise any dictionaries, but the |
| 284 // response claimed to be SDCH. There is no way to repair this situation: the |
| 285 // original request already didn't advertise any dictionaries, and retrying it |
| 286 // would likely have the/ same result. Blacklist the domain and try passing |
| 287 // through. |
| 288 if (!dictionary_set_) { |
| 289 sdch_manager_->BlacklistDomain(url_, SDCH_PASSING_THROUGH_NON_SDCH); |
| 290 LogCorruptionDetection(net_log_, is_cached_content_, RESPONSE_ENCODING_LIE); |
| 291 return PASS_THROUGH; |
| 292 } |
| 293 // Since SDCH dictionaries are advertised, this is a corrupt SDCH response. |
| 294 LogCorruptionDetection(net_log_, is_cached_content_, RESPONSE_CORRUPT_SDCH); |
| 295 return IssueMetaRefreshIfPossible(replace_output); |
| 296 } |
| 297 |
| 298 // Dictionary fails to load when we have a plausible dictionay id. There are |
| 299 // many possible causes: a cached SDCH-ified reply for which the SdchManager did |
| 300 // not have the dictionary or a corrupted response. These are handled here by |
| 301 // issuing a meta-refresh except the case where response code is 404. |
| 302 SdchPolicyDelegate::ErrorRecovery SdchPolicyDelegate::OnGetDictionaryError( |
| 303 std::string* replace_output) { |
| 304 if (possible_pass_through_) { |
| 305 LogCorruptionDetection(net_log_, is_cached_content_, |
| 306 RESPONSE_TENTATIVE_SDCH); |
| 307 // Ideally we should return PASS_THROUGH here, but this is done to match |
| 308 // the old behavior in sdch_filter.cc. |
| 309 } |
| 310 // HTTP 404 might be an unencoded error page, so if decoding failed, pass it |
| 311 // through. TODO(xunjieli): Remove this case crbug.com/516773. |
| 312 if (response_code_ == 404) { |
| 313 SdchManager::LogSdchProblem(net_log_, SDCH_PASS_THROUGH_404_CODE); |
| 314 LogCorruptionDetection(net_log_, is_cached_content_, RESPONSE_404); |
| 315 return PASS_THROUGH; |
| 316 } |
| 317 SdchManager::LogSdchProblem(net_log_, SDCH_DICTIONARY_HASH_NOT_FOUND); |
| 318 LogCorruptionDetection(net_log_, is_cached_content_, RESPONSE_NO_DICTIONARY); |
| 319 return IssueMetaRefreshIfPossible(replace_output); |
| 320 } |
| 321 |
| 322 SdchPolicyDelegate::ErrorRecovery SdchPolicyDelegate::OnDecodingError( |
| 323 std::string* replace_output) { |
| 324 // A decoding error, as opposed to a dictionary error, indicates a |
| 325 // decompression failure partway through the payload of the SDCH stream, |
| 326 // which means that the filter already witnessed a valid dictionary ID and |
| 327 // successfully retrieved a dictionary for it. Decoding errors are not |
| 328 // recoverable and it is not appropriate to stop decoding, so there are |
| 329 // relatively few error cases here. |
| 330 // |
| 331 // In particular, a decoding error for an HTML payload is recoverable by |
| 332 // issuing a meta-refresh, but to avoid having that happen too often, this |
| 333 // class also temporarily blacklists the domain. A decoding error for a |
| 334 // non-HTML payload is unrecoverable, so such an error gets a permanent |
| 335 // blacklist entry. If the content was cached, no blacklisting is needed. |
| 336 // TODO(xunjieli): This case should be removed. crbug.com/651821. |
| 337 return IssueMetaRefreshIfPossible(replace_output); |
| 338 } |
| 339 |
| 340 bool SdchPolicyDelegate::OnGetDictionary(const std::string& server_id, |
| 341 const std::string** text) { |
| 342 if (dictionary_set_) { |
| 343 *text = dictionary_set_->GetDictionaryText(server_id); |
| 344 if (*text) { |
| 345 server_id_ = server_id; |
| 346 return true; |
| 347 } |
| 348 } |
| 349 // This is a hack. Naively, the dictionaries available for |
| 350 // decoding should be only the ones advertised. However, there are |
| 351 // cases, specifically resources encoded with old dictionaries living |
| 352 // in the cache, that mean the full set of dictionaries should be made |
| 353 // available for decoding. It's not known how often this happens; |
| 354 // if it happens rarely enough, this code can be removed. |
| 355 // |
| 356 // TODO(rdsmith): Long-term, a better solution is necessary, since |
| 357 // an entry in the cache being encoded with the dictionary doesn't |
| 358 // guarantee that the dictionary is present. That solution probably |
| 359 // involves storing unencoded resources in the cache, but might |
| 360 // involve evicting encoded resources on dictionary removal. |
| 361 // See http://crbug.com/383405. |
| 362 SdchProblemCode rv = SDCH_OK; |
| 363 unexpected_dictionary_set_ = |
| 364 sdch_manager_->GetDictionarySetByHash(url_, server_id, &rv); |
| 365 if (unexpected_dictionary_set_) { |
| 366 *text = unexpected_dictionary_set_->GetDictionaryText(server_id); |
| 367 SdchManager::LogSdchProblem( |
| 368 net_log_, is_cached_content_ ? SDCH_UNADVERTISED_DICTIONARY_USED_CACHED |
| 369 : SDCH_UNADVERTISED_DICTIONARY_USED); |
| 370 if (*text) { |
| 371 server_id_ = server_id; |
| 372 return true; |
| 373 } |
| 374 } |
| 375 return false; |
| 376 } |
| 377 |
| 378 void SdchPolicyDelegate::OnStreamDestroyed( |
| 379 SdchSourceStream::InputState input_state, |
| 380 bool buffered_output_present, |
| 381 bool decoding_not_finished) { |
| 382 if (decoding_not_finished) { |
| 383 SdchManager::LogSdchProblem(net_log_, SDCH_INCOMPLETE_SDCH_CONTENT); |
| 384 // Make it possible for the user to hit reload, and get non-sdch content. |
| 385 // Note this will "wear off" quickly enough, and is just meant to assure |
| 386 // in some rare case that the user is not stuck. |
| 387 sdch_manager_->BlacklistDomain(url_, SDCH_INCOMPLETE_SDCH_CONTENT); |
| 388 } |
| 389 // Filter chaining error, or premature teardown. |
| 390 if (buffered_output_present) |
| 391 SdchManager::LogSdchProblem(net_log_, SDCH_UNFLUSHED_CONTENT); |
| 392 |
| 393 // Only record packet stats for non-cached content. |
| 394 if (is_cached_content_) { |
| 395 // Not a real error, but it is useful to have this tally. |
| 396 // TODO(jar): Remove this stat after SDCH stability is validated. |
| 397 SdchManager::LogSdchProblem(net_log_, SDCH_CACHE_DECODED); |
| 398 return; // We don't need timing stats, and we aready got ratios. |
| 399 } |
| 400 switch (input_state) { |
| 401 case SdchSourceStream::STATE_DECODE: { |
| 402 job_->RecordPacketStats(FilterContext::StatisticSelector::SDCH_DECODE); |
| 403 // Allow latency experiments to proceed. |
| 404 sdch_manager_->SetAllowLatencyExperiment(url_, true); |
| 405 |
| 406 // Notify successful dictionary usage. |
| 407 DCHECK(!server_id_.empty()); |
| 408 sdch_manager_->OnDictionaryUsed(server_id_); |
| 409 return; |
| 410 } |
| 411 case SdchSourceStream::STATE_LOAD_DICTIONARY: |
| 412 SdchManager::LogSdchProblem(net_log_, SDCH_PRIOR_TO_DICTIONARY); |
| 413 return; |
| 414 case SdchSourceStream::STATE_PASS_THROUGH: |
| 415 job_->RecordPacketStats( |
| 416 FilterContext::StatisticSelector::SDCH_PASSTHROUGH); |
| 417 return; |
| 418 case SdchSourceStream::STATE_OUTPUT_REPLACE: |
| 419 // This is meta refresh case. Already accounted for when set. |
| 420 return; |
| 421 } // end of switch. |
| 422 } |
| 423 |
| 424 // TODO(xunjieli): Remove meta refresh. crbug.com/651821. |
| 425 SdchPolicyDelegate::ErrorRecovery |
| 426 SdchPolicyDelegate::IssueMetaRefreshIfPossible(std::string* replace_output) { |
| 427 // Errors for non-HTML payloads are unrecoverable and get the domain |
| 428 // blacklisted indefinitely. |
| 429 if (mime_type_.npos == mime_type_.find("text/html")) { |
| 430 SdchProblemCode problem = |
| 431 (is_cached_content_ ? SDCH_CACHED_META_REFRESH_UNSUPPORTED |
| 432 : SDCH_META_REFRESH_UNSUPPORTED); |
| 433 sdch_manager_->BlacklistDomainForever(url_, problem); |
| 434 SdchManager::LogSdchProblem(net_log_, problem); |
| 435 return NONE; |
| 436 } |
| 437 |
| 438 if (is_cached_content_) { |
| 439 // Cached content is a probably startup tab, so just get the fresh content |
| 440 // and try again, without disabling SDCH. |
| 441 SdchManager::LogSdchProblem(net_log_, SDCH_META_REFRESH_CACHED_RECOVERY); |
| 442 } else { |
| 443 // Since it wasn't in the cache, blacklist for some period to get the |
| 444 // correct content. |
| 445 sdch_manager_->BlacklistDomain(url_, SDCH_META_REFRESH_RECOVERY); |
| 446 SdchManager::LogSdchProblem(net_log_, SDCH_META_REFRESH_RECOVERY); |
| 447 } |
| 448 |
| 449 *replace_output = std::string(kRefreshHtml, strlen(kRefreshHtml)); |
| 450 return REPLACE_OUTPUT; |
| 451 } |
| 452 |
| 453 } // namespace net |
OLD | NEW |