thgraphmaster.cpp 103 KB

1234567891011121314151617181920212223242526272829303132333435363738394041424344454647484950515253545556575859606162636465666768697071727374757677787980818283848586878889909192939495969798991001011021031041051061071081091101111121131141151161171181191201211221231241251261271281291301311321331341351361371381391401411421431441451461471481491501511521531541551561571581591601611621631641651661671681691701711721731741751761771781791801811821831841851861871881891901911921931941951961971981992002012022032042052062072082092102112122132142152162172182192202212222232242252262272282292302312322332342352362372382392402412422432442452462472482492502512522532542552562572582592602612622632642652662672682692702712722732742752762772782792802812822832842852862872882892902912922932942952962972982993003013023033043053063073083093103113123133143153163173183193203213223233243253263273283293303313323333343353363373383393403413423433443453463473483493503513523533543553563573583593603613623633643653663673683693703713723733743753763773783793803813823833843853863873883893903913923933943953963973983994004014024034044054064074084094104114124134144154164174184194204214224234244254264274284294304314324334344354364374384394404414424434444454464474484494504514524534544554564574584594604614624634644654664674684694704714724734744754764774784794804814824834844854864874884894904914924934944954964974984995005015025035045055065075085095105115125135145155165175185195205215225235245255265275285295305315325335345355365375385395405415425435445455465475485495505515525535545555565575585595605615625635645655665675685695705715725735745755765775785795805815825835845855865875885895905915925935945955965975985996006016026036046056066076086096106116126136146156166176186196206216226236246256266276286296306316326336346356366376386396406416426436446456466476486496506516526536546556566576586596606616626636646656666676686696706716726736746756766776786796806816826836846856866876886896906916926936946956966976986997007017027037047057067077087097107117127137147157167177187197207217227237247257267277287297307317327337347357367377387397407417427437447457467477487497507517527537547557567577587597607617627637647657667677687697707717727737747757767777787797807817827837847857867877887897907917927937947957967977987998008018028038048058068078088098108118128138148158168178188198208218228238248258268278288298308318328338348358368378388398408418428438448458468478488498508518528538548558568578588598608618628638648658668678688698708718728738748758768778788798808818828838848858868878888898908918928938948958968978988999009019029039049059069079089099109119129139149159169179189199209219229239249259269279289299309319329339349359369379389399409419429439449459469479489499509519529539549559569579589599609619629639649659669679689699709719729739749759769779789799809819829839849859869879889899909919929939949959969979989991000100110021003100410051006100710081009101010111012101310141015101610171018101910201021102210231024102510261027102810291030103110321033103410351036103710381039104010411042104310441045104610471048104910501051105210531054105510561057105810591060106110621063106410651066106710681069107010711072107310741075107610771078107910801081108210831084108510861087108810891090109110921093109410951096109710981099110011011102110311041105110611071108110911101111111211131114111511161117111811191120112111221123112411251126112711281129113011311132113311341135113611371138113911401141114211431144114511461147114811491150115111521153115411551156115711581159116011611162116311641165116611671168116911701171117211731174117511761177117811791180118111821183118411851186118711881189119011911192119311941195119611971198119912001201120212031204120512061207120812091210121112121213121412151216121712181219122012211222122312241225122612271228122912301231123212331234123512361237123812391240124112421243124412451246124712481249125012511252125312541255125612571258125912601261126212631264126512661267126812691270127112721273127412751276127712781279128012811282128312841285128612871288128912901291129212931294129512961297129812991300130113021303130413051306130713081309131013111312131313141315131613171318131913201321132213231324132513261327132813291330133113321333133413351336133713381339134013411342134313441345134613471348134913501351135213531354135513561357135813591360136113621363136413651366136713681369137013711372137313741375137613771378137913801381138213831384138513861387138813891390139113921393139413951396139713981399140014011402140314041405140614071408140914101411141214131414141514161417141814191420142114221423142414251426142714281429143014311432143314341435143614371438143914401441144214431444144514461447144814491450145114521453145414551456145714581459146014611462146314641465146614671468146914701471147214731474147514761477147814791480148114821483148414851486148714881489149014911492149314941495149614971498149915001501150215031504150515061507150815091510151115121513151415151516151715181519152015211522152315241525152615271528152915301531153215331534153515361537153815391540154115421543154415451546154715481549155015511552155315541555155615571558155915601561156215631564156515661567156815691570157115721573157415751576157715781579158015811582158315841585158615871588158915901591159215931594159515961597159815991600160116021603160416051606160716081609161016111612161316141615161616171618161916201621162216231624162516261627162816291630163116321633163416351636163716381639164016411642164316441645164616471648164916501651165216531654165516561657165816591660166116621663166416651666166716681669167016711672167316741675167616771678167916801681168216831684168516861687168816891690169116921693169416951696169716981699170017011702170317041705170617071708170917101711171217131714171517161717171817191720172117221723172417251726172717281729173017311732173317341735173617371738173917401741174217431744174517461747174817491750175117521753175417551756175717581759176017611762176317641765176617671768176917701771177217731774177517761777177817791780178117821783178417851786178717881789179017911792179317941795179617971798179918001801180218031804180518061807180818091810181118121813181418151816181718181819182018211822182318241825182618271828182918301831183218331834183518361837183818391840184118421843184418451846184718481849185018511852185318541855185618571858185918601861186218631864186518661867186818691870187118721873187418751876187718781879188018811882188318841885188618871888188918901891189218931894189518961897189818991900190119021903190419051906190719081909191019111912191319141915191619171918191919201921192219231924192519261927192819291930193119321933193419351936193719381939194019411942194319441945194619471948194919501951195219531954195519561957195819591960196119621963196419651966196719681969197019711972197319741975197619771978197919801981198219831984198519861987198819891990199119921993199419951996199719981999200020012002200320042005200620072008200920102011201220132014201520162017201820192020202120222023202420252026202720282029203020312032203320342035203620372038203920402041204220432044204520462047204820492050205120522053205420552056205720582059206020612062206320642065206620672068206920702071207220732074207520762077207820792080208120822083208420852086208720882089209020912092209320942095209620972098209921002101210221032104210521062107210821092110211121122113211421152116211721182119212021212122212321242125212621272128212921302131213221332134213521362137213821392140214121422143214421452146214721482149215021512152215321542155215621572158215921602161216221632164216521662167216821692170217121722173217421752176217721782179218021812182218321842185218621872188218921902191219221932194219521962197219821992200220122022203220422052206220722082209221022112212221322142215221622172218221922202221222222232224222522262227222822292230223122322233223422352236223722382239224022412242224322442245224622472248224922502251225222532254225522562257225822592260226122622263226422652266226722682269227022712272227322742275227622772278227922802281228222832284228522862287228822892290229122922293229422952296229722982299230023012302230323042305230623072308230923102311231223132314231523162317231823192320232123222323232423252326232723282329233023312332233323342335233623372338233923402341234223432344234523462347234823492350235123522353235423552356235723582359236023612362236323642365236623672368236923702371237223732374237523762377237823792380238123822383238423852386238723882389239023912392239323942395239623972398239924002401240224032404240524062407240824092410241124122413241424152416241724182419242024212422242324242425242624272428242924302431243224332434243524362437243824392440244124422443244424452446244724482449245024512452245324542455245624572458245924602461246224632464246524662467246824692470247124722473247424752476247724782479248024812482248324842485248624872488248924902491249224932494249524962497249824992500250125022503250425052506250725082509251025112512251325142515251625172518251925202521252225232524252525262527252825292530253125322533253425352536253725382539254025412542254325442545254625472548254925502551255225532554255525562557255825592560256125622563256425652566256725682569257025712572257325742575257625772578257925802581258225832584258525862587258825892590259125922593259425952596259725982599260026012602260326042605260626072608260926102611261226132614261526162617261826192620262126222623262426252626262726282629263026312632263326342635263626372638263926402641264226432644264526462647264826492650265126522653265426552656265726582659266026612662266326642665266626672668266926702671267226732674267526762677267826792680268126822683268426852686268726882689269026912692269326942695269626972698269927002701270227032704270527062707270827092710271127122713271427152716271727182719272027212722272327242725272627272728272927302731273227332734273527362737273827392740274127422743274427452746274727482749275027512752275327542755275627572758275927602761276227632764276527662767276827692770277127722773277427752776277727782779278027812782278327842785278627872788278927902791279227932794279527962797279827992800280128022803280428052806280728082809281028112812281328142815281628172818281928202821282228232824282528262827282828292830283128322833283428352836283728382839284028412842284328442845284628472848284928502851285228532854285528562857285828592860286128622863286428652866286728682869287028712872287328742875287628772878287928802881288228832884288528862887288828892890289128922893289428952896289728982899290029012902290329042905290629072908290929102911291229132914291529162917291829192920292129222923292429252926292729282929293029312932293329342935293629372938293929402941294229432944294529462947294829492950295129522953295429552956295729582959296029612962296329642965296629672968296929702971297229732974297529762977297829792980298129822983298429852986298729882989299029912992299329942995299629972998299930003001300230033004300530063007300830093010301130123013301430153016301730183019
  1. /*##############################################################################
  2. HPCC SYSTEMS software Copyright (C) 2012 HPCC Systems®.
  3. Licensed under the Apache License, Version 2.0 (the "License");
  4. you may not use this file except in compliance with the License.
  5. You may obtain a copy of the License at
  6. http://www.apache.org/licenses/LICENSE-2.0
  7. Unless required by applicable law or agreed to in writing, software
  8. distributed under the License is distributed on an "AS IS" BASIS,
  9. WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
  10. See the License for the specific language governing permissions and
  11. limitations under the License.
  12. ############################################################################## */
  13. #include <limits.h>
  14. #include <stdlib.h>
  15. #include "jprop.hpp"
  16. #include "jexcept.hpp"
  17. #include "jiter.ipp"
  18. #include "jlzw.hpp"
  19. #include "jsocket.hpp"
  20. #include "jset.hpp"
  21. #include "jsort.hpp"
  22. #include "portlist.h"
  23. #include "jhtree.hpp"
  24. #include "mputil.hpp"
  25. #include "dllserver.hpp"
  26. #include "dautils.hpp"
  27. #include "danqs.hpp"
  28. #include "daclient.hpp"
  29. #include "daaudit.hpp"
  30. #include "wujobq.hpp"
  31. #include "thorport.hpp"
  32. #include "commonext.hpp"
  33. #include "thorxmlread.hpp"
  34. #include "thorplugin.hpp"
  35. #include "thormisc.hpp"
  36. #include "thgraphmaster.ipp"
  37. #include "thdemonserver.hpp"
  38. #include "rtlds_imp.hpp"
  39. #include "eclhelper.hpp"
  40. #include "thexception.hpp"
  41. #include "thactivitymaster.ipp"
  42. #include "thmem.hpp"
  43. #include "thcompressutil.hpp"
  44. using roxiemem::OwnedRoxieString;
  45. static CriticalSection *jobManagerCrit;
  46. MODULE_INIT(INIT_PRIORITY_STANDARD)
  47. {
  48. jobManagerCrit = new CriticalSection;
  49. return true;
  50. }
  51. MODULE_EXIT()
  52. {
  53. delete jobManagerCrit;
  54. }
  55. unsigned uniqGraphId = 1;
  56. #define FATAL_TIMEOUT 60
  57. class CFatalHandler : public CTimeoutTrigger, implements IFatalHandler
  58. {
  59. public:
  60. IMPLEMENT_IINTERFACE;
  61. CFatalHandler(unsigned timeout) : CTimeoutTrigger(timeout, "EXCEPTION")
  62. {
  63. }
  64. virtual bool action()
  65. {
  66. StringBuffer s("FAILED TO RECOVER FROM EXCEPTION, STOPPING THOR");
  67. FLLOG(MCoperatorWarning, thorJob, exception, s.str());
  68. Owned<IJobManager> jobManager = getJobManager();
  69. if (jobManager)
  70. {
  71. jobManager->fatal(exception);
  72. jobManager.clear();
  73. }
  74. return true;
  75. }
  76. // IFatalHandler
  77. virtual void inform(IException *e)
  78. {
  79. CTimeoutTrigger::inform(e);
  80. }
  81. virtual void clear()
  82. {
  83. CTimeoutTrigger::clear();
  84. }
  85. };
  86. /////
  87. CSlaveMessageHandler::CSlaveMessageHandler(CJobMaster &_job, mptag_t _mptag) : threaded("CSlaveMessageHandler"), job(_job), mptag(_mptag)
  88. {
  89. stopped = false;
  90. threaded.init(this);
  91. childGraphInitTimeout = job.getOptUInt(THOROPT_CHILD_GRAPH_INIT_TIMEOUT, 5*60) * 1000; // default 5 minutes
  92. }
  93. CSlaveMessageHandler::~CSlaveMessageHandler()
  94. {
  95. stop();
  96. }
  97. void CSlaveMessageHandler::stop()
  98. {
  99. if (!stopped)
  100. {
  101. stopped = true;
  102. job.queryJobChannel(0).queryJobComm().cancel(0, mptag);
  103. threaded.join();
  104. }
  105. }
  106. void CSlaveMessageHandler::threadmain()
  107. {
  108. try
  109. {
  110. for (;;)
  111. {
  112. rank_t sender;
  113. CMessageBuffer msg;
  114. if (stopped || !job.queryJobChannel(0).queryJobComm().recv(msg, RANK_ALL, mptag, &sender))
  115. break;
  116. SlaveMsgTypes msgType;
  117. msg.read((int &)msgType);
  118. switch (msgType)
  119. {
  120. case smt_errorMsg:
  121. {
  122. unsigned slave;
  123. msg.read(slave);
  124. Owned<IThorException> e = deserializeThorException(msg);
  125. e->setSlave(slave+1);
  126. Owned<CGraphBase> graph = job.queryJobChannel(0).getGraph(e->queryGraphId());
  127. if (graph)
  128. {
  129. activity_id id = e->queryActivityId();
  130. if (id)
  131. {
  132. CGraphElementBase *elem = graph->queryElement(id);
  133. CActivityBase *act = elem->queryActivity();
  134. if (act)
  135. act->fireException(e);
  136. else
  137. graph->fireException(e);
  138. }
  139. else
  140. graph->fireException(e);
  141. }
  142. else
  143. job.fireException(e);
  144. if (msg.getReplyTag() <= TAG_REPLY_BASE)
  145. {
  146. msg.clear();
  147. job.queryJobChannel(0).queryJobComm().reply(msg);
  148. }
  149. break;
  150. }
  151. case smt_dataReq:
  152. {
  153. graph_id gid;
  154. activity_id id;
  155. unsigned slave;
  156. msg.read(slave);
  157. msg.read(gid);
  158. msg.read(id);
  159. msg.clear();
  160. Owned<CGraphBase> graph = job.queryJobChannel(0).getGraph(gid);
  161. if (graph)
  162. {
  163. CMasterGraphElement *e = (CMasterGraphElement *)graph->queryElement(id);
  164. e->queryActivity()->getInitializationData(slave, msg);
  165. }
  166. job.queryJobChannel(0).queryJobComm().reply(msg);
  167. break;
  168. }
  169. case smt_initGraphReq:
  170. {
  171. graph_id gid;
  172. msg.read(gid);
  173. Owned<CMasterGraph> graph = (CMasterGraph *)job.queryJobChannel(0).getGraph(gid);
  174. assertex(graph);
  175. {
  176. CriticalBlock b(graph->queryCreateLock());
  177. Owned<IThorActivityIterator> iter = graph->getIterator();
  178. // onCreate all
  179. ForEach (*iter)
  180. {
  181. CMasterGraphElement &element = (CMasterGraphElement &)iter->query();
  182. element.onCreate();
  183. }
  184. }
  185. msg.clear();
  186. graph->serializeCreateContexts(msg);
  187. job.queryJobChannel(0).queryJobComm().reply(msg);
  188. break;
  189. }
  190. case smt_initActDataReq:
  191. {
  192. graph_id gid;
  193. msg.read(gid);
  194. unsigned slave;
  195. msg.read(slave);
  196. Owned<CMasterGraph> graph = (CMasterGraph *)job.queryJobChannel(0).getGraph(gid);
  197. assertex(graph);
  198. CGraphElementArray toSerialize;
  199. CriticalBlock b(graph->queryCreateLock());
  200. size32_t parentExtractSz;
  201. msg.read(parentExtractSz);
  202. const byte *parentExtract = NULL;
  203. if (parentExtractSz)
  204. {
  205. parentExtract = msg.readDirect(parentExtractSz);
  206. StringBuffer msg("Graph(");
  207. msg.append(graph->queryGraphId()).append(") - initializing master graph with parentExtract ").append(parentExtractSz).append(" bytes");
  208. DBGLOG("%s", msg.str());
  209. parentExtract = graph->setParentCtx(parentExtractSz, parentExtract);
  210. }
  211. Owned<IException> exception;
  212. for (;;)
  213. {
  214. activity_id id;
  215. msg.read(id);
  216. if (!id)
  217. break;
  218. CMasterGraphElement *element = (CMasterGraphElement *)graph->queryElement(id);
  219. assertex(element);
  220. try
  221. {
  222. element->reset();
  223. size32_t startCtxLen;
  224. msg.read(startCtxLen);
  225. element->doCreateActivity(parentExtractSz, parentExtract, startCtxLen ? &msg : nullptr);
  226. if (element->queryActivity())
  227. element->preStart(parentExtractSz, parentExtract);
  228. }
  229. catch (IException *e)
  230. {
  231. EXCLOG(e, NULL);
  232. exception.setown(e);
  233. break;
  234. }
  235. CActivityBase *activity = element->queryActivity();
  236. if (activity)
  237. element->sentActInitData->set(slave, 0); // clear to permit serializeActivityInitData to resend
  238. toSerialize.append(*LINK(element));
  239. }
  240. msg.clear();
  241. mptag_t replyTag = job.queryJobChannel(0).queryMPServer().createReplyTag();
  242. msg.append(replyTag); // second reply
  243. if (exception)
  244. {
  245. msg.append(true);
  246. serializeException(exception, msg);
  247. }
  248. else
  249. {
  250. msg.append(false);
  251. CGraphElementArrayIterator iter(toSerialize);
  252. graph->serializeActivityInitData(slave, msg, iter);
  253. }
  254. job.queryJobChannel(0).queryJobComm().reply(msg);
  255. if (!job.queryJobChannel(0).queryJobComm().recv(msg, slave+1, replyTag, NULL, childGraphInitTimeout))
  256. throwUnexpected();
  257. if (exception)
  258. throw exception.getClear();
  259. bool error;
  260. msg.read(error);
  261. if (error)
  262. {
  263. Owned<IThorException> e = deserializeThorException(msg);
  264. e->setSlave(slave);
  265. StringBuffer tmpStr("Slave ");
  266. job.queryJobGroup().queryNode(slave).endpoint().getUrlStr(tmpStr);
  267. GraphPrintLog(graph, e, "%s", tmpStr.append(": slave initialization error").str());
  268. throw e.getClear();
  269. }
  270. break;
  271. }
  272. case smt_getPhysicalName:
  273. {
  274. LOG(MCdebugProgress, unknownJob, "getPhysicalName called from node %d", sender-1);
  275. StringAttr logicalName;
  276. unsigned partNo;
  277. bool create;
  278. msg.read(logicalName);
  279. msg.read(partNo);
  280. msg.read(create);
  281. msg.clear();
  282. StringBuffer phys;
  283. if (create && !job.queryCreatedFile(logicalName)) // not sure who would do this ever??
  284. queryThorFileManager().getPublishPhysicalName(job, logicalName, partNo, phys);
  285. else
  286. queryThorFileManager().getPhysicalName(job, logicalName, partNo, phys);
  287. msg.append(phys);
  288. job.queryJobChannel(0).queryJobComm().reply(msg);
  289. break;
  290. }
  291. case smt_getFileOffset:
  292. {
  293. LOG(MCdebugProgress, unknownJob, "getFileOffset called from node %d", sender-1);
  294. StringAttr logicalName;
  295. unsigned partNo;
  296. msg.read(logicalName);
  297. msg.read(partNo);
  298. msg.clear();
  299. offset_t offset = queryThorFileManager().getFileOffset(job, logicalName, partNo);
  300. msg.append(offset);
  301. job.queryJobChannel(0).queryJobComm().reply(msg);
  302. break;
  303. }
  304. case smt_actMsg:
  305. {
  306. LOG(MCdebugProgress, unknownJob, "smt_actMsg called from node %d", sender-1);
  307. graph_id gid;
  308. msg.read(gid);
  309. activity_id id;
  310. msg.read(id);
  311. Owned<CMasterGraph> graph = (CMasterGraph *)job.queryJobChannel(0).getGraph(gid);
  312. assertex(graph);
  313. CMasterGraphElement *container = (CMasterGraphElement *)graph->queryElement(id);
  314. assertex(container);
  315. CMasterActivity *activity = (CMasterActivity *)container->queryActivity();
  316. assertex(activity);
  317. activity->handleSlaveMessage(msg); // don't block
  318. break;
  319. }
  320. case smt_getresult:
  321. {
  322. unsigned slave;
  323. msg.read(slave);
  324. LOG(MCdebugProgress, unknownJob, "smt_getresult called from slave %d", slave);
  325. graph_id gid;
  326. msg.read(gid);
  327. activity_id ownerId;
  328. msg.read(ownerId);
  329. unsigned resultId;
  330. msg.read(resultId);
  331. CJobChannel &jobChannel = job.queryJobChannel(0);
  332. mptag_t replyTag = jobChannel.deserializeMPTag(msg);
  333. Owned<IThorResult> result = jobChannel.getOwnedResult(gid, ownerId, resultId);
  334. Owned<IRowStream> resultStream = result->getRowStream();
  335. sendInChunks(job.queryJobChannel(0).queryJobComm(), slave+1, replyTag, resultStream, result->queryRowInterfaces());
  336. break;
  337. }
  338. }
  339. }
  340. }
  341. catch (IException *e)
  342. {
  343. job.fireException(e);
  344. e->Release();
  345. }
  346. }
  347. //////////////////////
  348. CMasterActivity::CMasterActivity(CGraphElementBase *_container) : CActivityBase(_container), threaded("CMasterActivity", this), timingInfo(_container->queryJob())
  349. {
  350. notedWarnings = createThreadSafeBitSet();
  351. mpTag = TAG_NULL;
  352. data = new MemoryBuffer[container.queryJob().querySlaves()];
  353. asyncStart = false;
  354. if (container.isSink())
  355. progressInfo.append(*new ProgressInfo(queryJob()));
  356. else
  357. {
  358. unsigned o=0;
  359. for (; o<container.getOutputs(); o++)
  360. progressInfo.append(*new ProgressInfo(queryJob()));
  361. }
  362. }
  363. CMasterActivity::~CMasterActivity()
  364. {
  365. if (asyncStart)
  366. threaded.join();
  367. notedWarnings->Release();
  368. queryJob().freeMPTag(mpTag);
  369. delete [] data;
  370. }
  371. void CMasterActivity::addReadFile(IDistributedFile *file, bool temp)
  372. {
  373. readFiles.append(*LINK(file));
  374. if (!temp) // NB: Temps not listed in workunit
  375. queryThorFileManager().noteFileRead(container.queryJob(), file);
  376. }
  377. IDistributedFile *CMasterActivity::queryReadFile(unsigned f)
  378. {
  379. if (f>=readFiles.ordinality())
  380. return NULL;
  381. return &readFiles.item(f);
  382. }
  383. void CMasterActivity::preStart(size32_t parentExtractSz, const byte *parentExtract)
  384. {
  385. CActivityBase::preStart(parentExtractSz, parentExtract);
  386. IArrayOf<IDistributedFile> tmpFiles;
  387. tmpFiles.swapWith(readFiles);
  388. ForEachItemIn(f, tmpFiles)
  389. {
  390. IDistributedFile &file = tmpFiles.item(f);
  391. IDistributedSuperFile *super = file.querySuperFile();
  392. if (super)
  393. getSuperFileSubs(super, readFiles, true);
  394. else
  395. readFiles.append(*LINK(&file));
  396. }
  397. }
  398. MemoryBuffer &CMasterActivity::queryInitializationData(unsigned slave) const
  399. { // NB: not intended to be called by multiple threads.
  400. return data[slave].reset();
  401. }
  402. MemoryBuffer &CMasterActivity::getInitializationData(unsigned slave, MemoryBuffer &dst) const
  403. {
  404. return dst.append(data[slave]);
  405. }
  406. void CMasterActivity::threadmain()
  407. {
  408. try
  409. {
  410. process();
  411. }
  412. catch (IException *e)
  413. {
  414. Owned<IException> e2;
  415. if (QUERYINTERFACE(e, ISEH_Exception))
  416. e2.setown(MakeThorFatal(e, TE_SEH, "(SEH)"));
  417. else
  418. e2.setown(MakeActivityException(this, e, "Master exception"));
  419. e->Release();
  420. ActPrintLog(e2, "In CMasterActivity::threadmain");
  421. fireException(e2);
  422. }
  423. catch (CATCHALL)
  424. {
  425. Owned<IException> e = MakeThorFatal(NULL, TE_MasterProcessError, "FATAL: Unknown master process exception kind=%s, id=%" ACTPF "d", activityKindStr(container.getKind()), container.queryId());
  426. ActPrintLog(e, "In CMasterActivity::threadmain");
  427. fireException(e);
  428. }
  429. }
  430. void CMasterActivity::init()
  431. {
  432. readFiles.kill();
  433. }
  434. void CMasterActivity::startProcess(bool async)
  435. {
  436. if (async)
  437. {
  438. asyncStart = true;
  439. threaded.start();
  440. }
  441. else
  442. threadmain();
  443. }
  444. bool CMasterActivity::wait(unsigned timeout)
  445. {
  446. if (!asyncStart)
  447. return true;
  448. return threaded.join(timeout);
  449. }
  450. void CMasterActivity::kill()
  451. {
  452. CActivityBase::kill();
  453. readFiles.kill();
  454. }
  455. bool CMasterActivity::fireException(IException *_e)
  456. {
  457. IThorException *e = QUERYINTERFACE(_e, IThorException);
  458. if (!e) return false;
  459. switch (e->errorCode())
  460. {
  461. case TE_LargeBufferWarning:
  462. case TE_MoxieIndarOverflow:
  463. case TE_BuildIndexFewExcess:
  464. case TE_FetchMisaligned:
  465. case TE_FetchOutOfRange:
  466. case TE_CouldNotCreateLookAhead:
  467. case TE_SpillAdded:
  468. case TE_ReadPartialFromPipe:
  469. case TE_LargeAggregateTable:
  470. case TE_RemoteReadFailure:
  471. {
  472. if (!notedWarnings->testSet(e->errorCode()))
  473. CActivityBase::fireException(e);
  474. return true;
  475. }
  476. }
  477. return container.queryOwner().fireException(e);
  478. }
  479. void CMasterActivity::reset()
  480. {
  481. asyncStart = false;
  482. CActivityBase::reset();
  483. }
  484. void CMasterActivity::deserializeStats(unsigned node, MemoryBuffer &mb)
  485. {
  486. CriticalBlock b(progressCrit); // don't think needed
  487. unsigned __int64 localTimeNs;
  488. mb.read(localTimeNs);
  489. timingInfo.set(node, localTimeNs);
  490. rowcount_t count;
  491. ForEachItemIn(p, progressInfo)
  492. {
  493. mb.read(count);
  494. progressInfo.item(p).set(node, count);
  495. }
  496. }
  497. void CMasterActivity::getActivityStats(IStatisticGatherer & stats)
  498. {
  499. timingInfo.getStats(stats);
  500. }
  501. void CMasterActivity::getEdgeStats(IStatisticGatherer & stats, unsigned idx)
  502. {
  503. CriticalBlock b(progressCrit);
  504. if (progressInfo.isItem(idx))
  505. progressInfo.item(idx).getStats(stats);
  506. }
  507. void CMasterActivity::done()
  508. {
  509. CActivityBase::done();
  510. ForEachItemIn(s, readFiles)
  511. {
  512. IDistributedFile &file = readFiles.item(s);
  513. file.setAccessed();
  514. }
  515. }
  516. //////////////////////
  517. // CMasterGraphElement impl.
  518. //
  519. CMasterGraphElement::CMasterGraphElement(CGraphBase &_owner, IPropertyTree &_xgmml) : CGraphElementBase(_owner, _xgmml)
  520. {
  521. }
  522. bool CMasterGraphElement::checkUpdate()
  523. {
  524. if (!onlyUpdateIfChanged)
  525. return false;
  526. if (!globals->getPropBool("@updateEnabled", true) || 0 != queryJob().getWorkUnitValueInt("disableUpdate", 0))
  527. return false;
  528. bool doCheckUpdate = false;
  529. OwnedRoxieString filename;
  530. unsigned eclCRC;
  531. unsigned __int64 totalCRC;
  532. bool temporary = false;
  533. switch (getKind())
  534. {
  535. case TAKindexwrite:
  536. {
  537. IHThorIndexWriteArg *helper = (IHThorIndexWriteArg *)queryHelper();
  538. doCheckUpdate = 0 != (helper->getFlags() & TIWupdate);
  539. filename.setown(helper->getFileName());
  540. helper->getUpdateCRCs(eclCRC, totalCRC);
  541. break;
  542. }
  543. case TAKdiskwrite:
  544. case TAKcsvwrite:
  545. case TAKxmlwrite:
  546. case TAKjsonwrite:
  547. case TAKspillwrite:
  548. {
  549. IHThorDiskWriteArg *helper = (IHThorDiskWriteArg *)queryHelper();
  550. doCheckUpdate = 0 != (helper->getFlags() & TDWupdate);
  551. filename.setown(helper->getFileName());
  552. helper->getUpdateCRCs(eclCRC, totalCRC);
  553. if (TAKdiskwrite == getKind())
  554. temporary = 0 != (helper->getFlags() & (TDXtemporary|TDXjobtemp));
  555. else if (TAKspillwrite == getKind())
  556. temporary = true;
  557. break;
  558. }
  559. }
  560. if (doCheckUpdate)
  561. {
  562. StringAttr lfn;
  563. Owned<IDistributedFile> file = queryThorFileManager().lookup(queryJob(), filename, temporary, true, false, defaultPrivilegedUser);
  564. if (file)
  565. {
  566. IPropertyTree &props = file->queryAttributes();
  567. if ((eclCRC == props.getPropInt("@eclCRC")) && (totalCRC == props.getPropInt64("@totalCRC")))
  568. {
  569. // so this needs pruning
  570. Owned<IThorException> e = MakeActivityWarning(this, TE_UpToDate, "output file = '%s' - is up to date - it will not be rebuilt", file->queryLogicalName());
  571. queryOwner().fireException(e);
  572. return true;
  573. }
  574. }
  575. }
  576. return false;
  577. }
  578. void CMasterGraphElement::initActivity()
  579. {
  580. CriticalBlock b(crit);
  581. if (!initialized)
  582. {
  583. initialized = true;
  584. ((CMasterActivity *)queryActivity())->init();
  585. }
  586. }
  587. void CMasterGraphElement::reset()
  588. {
  589. CGraphElementBase::reset();
  590. if (activity && activity->needReInit())
  591. initialized = false;
  592. }
  593. void CMasterGraphElement::doCreateActivity(size32_t parentExtractSz, const byte *parentExtract, MemoryBuffer *startCtx)
  594. {
  595. createActivity();
  596. onStart(parentExtractSz, parentExtract, startCtx);
  597. initActivity();
  598. }
  599. void CMasterGraphElement::slaveDone(size32_t slaveIdx, MemoryBuffer &mb)
  600. {
  601. ((CMasterActivity *)queryActivity())->slaveDone(slaveIdx, mb);
  602. }
  603. //////
  604. ///////
  605. class CBarrierMaster : public CInterface, implements IBarrier
  606. {
  607. mptag_t tag;
  608. Linked<ICommunicator> comm;
  609. bool receiving;
  610. public:
  611. IMPLEMENT_IINTERFACE;
  612. CBarrierMaster(ICommunicator &_comm, mptag_t _tag) : comm(&_comm), tag(_tag)
  613. {
  614. receiving = false;
  615. }
  616. virtual const mptag_t queryTag() const { return tag; }
  617. virtual bool wait(bool exception, unsigned timeout)
  618. {
  619. Owned<IException> e;
  620. CTimeMon tm(timeout);
  621. unsigned s=comm->queryGroup().ordinality()-1;
  622. bool aborted = false;
  623. CMessageBuffer msg;
  624. Owned<IBitSet> raisedSet = createThreadSafeBitSet();
  625. unsigned remaining = timeout;
  626. while (s--)
  627. {
  628. rank_t sender;
  629. msg.clear();
  630. if (INFINITE != timeout && tm.timedout(&remaining))
  631. {
  632. if (exception)
  633. throw createBarrierAbortException();
  634. else
  635. return false;
  636. }
  637. {
  638. BooleanOnOff onOff(receiving);
  639. if (!comm->recv(msg, RANK_ALL, tag, &sender, remaining))
  640. break;
  641. }
  642. msg.read(aborted);
  643. bool hasExcept;
  644. msg.read(hasExcept);
  645. if (hasExcept && !e.get())
  646. e.setown(deserializeException(msg));
  647. sender = sender - 1; // 0 = master
  648. if (raisedSet->testSet(sender, true) && !aborted)
  649. IWARNLOG("CBarrierMaster, raise barrier message on tag %d, already received from slave %d", tag, sender);
  650. if (aborted) break;
  651. }
  652. msg.clear();
  653. msg.append(aborted);
  654. if (e)
  655. {
  656. msg.append(true);
  657. serializeException(e, msg);
  658. }
  659. else
  660. msg.append(false);
  661. if (INFINITE != timeout && tm.timedout(&remaining))
  662. {
  663. if (exception)
  664. throw createBarrierAbortException();
  665. else
  666. return false;
  667. }
  668. if (!comm->send(msg, RANK_ALL_OTHER, tag, INFINITE != timeout ? remaining : LONGTIMEOUT))
  669. throw MakeStringException(0, "CBarrierMaster::wait - Timeout sending to slaves");
  670. if (aborted)
  671. {
  672. if (!exception)
  673. return false;
  674. if (e)
  675. throw e.getClear();
  676. else
  677. throw createBarrierAbortException();
  678. }
  679. return true;
  680. }
  681. virtual void cancel(IException *e)
  682. {
  683. if (receiving)
  684. comm->cancel(RANK_ALL, tag);
  685. CMessageBuffer msg;
  686. msg.append(true);
  687. if (e)
  688. {
  689. msg.append(true);
  690. serializeException(e, msg);
  691. }
  692. else
  693. msg.append(false);
  694. if (!comm->send(msg, RANK_ALL_OTHER, tag, LONGTIMEOUT))
  695. throw MakeStringException(0, "CBarrierMaster::cancel - Timeout sending to slaves");
  696. }
  697. };
  698. /////////////
  699. class CMasterGraphTempHandler : public CGraphTempHandler
  700. {
  701. public:
  702. CMasterGraphTempHandler(CJobBase &job, bool errorOnMissing) : CGraphTempHandler(job, errorOnMissing) { }
  703. virtual bool removeTemp(const char *name)
  704. {
  705. queryThorFileManager().clearCacheEntry(name);
  706. return true;
  707. }
  708. virtual void registerFile(const char *name, graph_id graphId, unsigned usageCount, bool temp, WUFileKind fileKind, StringArray *clusters)
  709. {
  710. if (!temp || job.queryUseCheckpoints())
  711. {
  712. Owned<IWorkUnit> wu = &job.queryWorkUnit().lock();
  713. wu->addFile(name, clusters, usageCount, fileKind, job.queryGraphName());
  714. }
  715. else
  716. CGraphTempHandler::registerFile(name, graphId, usageCount, temp, fileKind, clusters);
  717. }
  718. virtual void deregisterFile(const char *name, bool kept) // NB: only called for temp files
  719. {
  720. if (kept || job.queryUseCheckpoints())
  721. {
  722. Owned<IWorkUnit> wu = &job.queryWorkUnit().lock();
  723. wu->releaseFile(name);
  724. }
  725. else
  726. CGraphTempHandler::deregisterFile(name);
  727. }
  728. virtual void clearTemps()
  729. {
  730. try
  731. {
  732. if (!job.queryPausing()) // temps of completed workunit will have been preserved and want to keep
  733. {
  734. Owned<IWorkUnit> lwu = &job.queryWorkUnit().lock();
  735. lwu->deleteTempFiles(job.queryGraphName(), false, false);
  736. }
  737. }
  738. catch (IException *e)
  739. {
  740. EXCLOG(e, "Problem deleting temp files");
  741. e->Release();
  742. }
  743. CGraphTempHandler::clearTemps();
  744. }
  745. };
  746. static const char * getResultText(StringBuffer & s, const char * stepname, unsigned sequence)
  747. {
  748. switch ((int)sequence)
  749. {
  750. case -1: return s.append("STORED('").append(stepname).append("')");
  751. case -2: return s.append("PERSIST('").append(stepname).append("')");
  752. case -3: return s.append("global('").append(stepname).append("')");
  753. default:
  754. if (stepname)
  755. return s.append(stepname);
  756. return s.append('#').append(sequence);
  757. }
  758. }
  759. class CThorCodeContextMaster : public CThorCodeContextBase
  760. {
  761. Linked<IConstWorkUnit> workunit;
  762. Owned<IDistributedFileTransaction> superfiletransaction;
  763. IConstWUResult * getResult(const char * name, unsigned sequence)
  764. {
  765. return getWorkUnitResult(workunit, name, sequence);
  766. }
  767. #define PROTECTED_GETRESULT(STEPNAME, SEQUENCE, KIND, KINDTEXT, ACTION) \
  768. LOG(MCdebugProgress, unknownJob, "getResult%s(%s,%d)", KIND, STEPNAME?STEPNAME:"", SEQUENCE); \
  769. Owned<IConstWUResult> r = getResultForGet(STEPNAME, SEQUENCE); \
  770. try \
  771. { \
  772. ACTION \
  773. } \
  774. catch (IException * e) { \
  775. StringBuffer s, text; e->errorMessage(text); e->Release(); \
  776. throw MakeStringException(TE_FailedToRetrieveWorkunitValue, "result %s in workunit contains an invalid " KINDTEXT " value [%s]", getResultText(s, STEPNAME, SEQUENCE), text.str()); \
  777. } \
  778. catch (CATCHALL) { StringBuffer s; throw MakeStringException(TE_FailedToRetrieveWorkunitValue, "value %s in workunit contains an invalid " KINDTEXT " value", getResultText(s, STEPNAME, SEQUENCE)); }
  779. public:
  780. CThorCodeContextMaster(CJobChannel &jobChannel, IConstWorkUnit &_workunit, ILoadedDllEntry &querySo, IUserDescriptor &userDesc) : CThorCodeContextBase(jobChannel, querySo, userDesc), workunit(&_workunit)
  781. {
  782. }
  783. // ICodeContext
  784. virtual unsigned getGraphLoopCounter() const override { return 0; }
  785. virtual IDebuggableContext *queryDebugContext() const override { return nullptr; }
  786. virtual void setResultBool(const char *name, unsigned sequence, bool result) override
  787. {
  788. WorkunitUpdate w(&workunit->lock());
  789. Owned<IWUResult> r = updateWorkUnitResult(w, name, sequence);
  790. if (r)
  791. {
  792. r->setResultBool(result);
  793. r->setResultStatus(ResultStatusCalculated);
  794. }
  795. else
  796. throw MakeStringException(TE_UnexpectedParameters, "Unexpected parameters to setResultBool");
  797. }
  798. virtual void setResultData(const char *name, unsigned sequence, int len, const void *result) override
  799. {
  800. WorkunitUpdate w(&workunit->lock());
  801. Owned<IWUResult> r = updateWorkUnitResult(w, name, sequence);
  802. if (r)
  803. {
  804. r->setResultData(result, len);
  805. r->setResultStatus(ResultStatusCalculated);
  806. }
  807. else
  808. throw MakeStringException(TE_UnexpectedParameters, "Unexpected parameters to setResultData");
  809. }
  810. virtual void setResultDecimal(const char * name, unsigned sequence, int len, int precision, bool isSigned, const void *val) override
  811. {
  812. WorkunitUpdate w(&workunit->lock());
  813. Owned<IWUResult> r = updateWorkUnitResult(w, name, sequence);
  814. if (r)
  815. {
  816. r->setResultDecimal(val, len);
  817. r->setResultStatus(ResultStatusCalculated);
  818. }
  819. else
  820. throw MakeStringException(TE_UnexpectedParameters, "Unexpected parameters to setResultDecimal");
  821. }
  822. virtual void setResultInt(const char *name, unsigned sequence, __int64 result, unsigned size) override
  823. {
  824. WorkunitUpdate w(&workunit->lock());
  825. Owned<IWUResult> r = updateWorkUnitResult(w, name, sequence);
  826. if (r)
  827. {
  828. r->setResultInt(result);
  829. r->setResultStatus(ResultStatusCalculated);
  830. }
  831. else
  832. throw MakeStringException(TE_UnexpectedParameters, "Unexpected parameters to setResultInt");
  833. }
  834. virtual void setResultRaw(const char *name, unsigned sequence, int len, const void *result) override
  835. {
  836. WorkunitUpdate w(&workunit->lock());
  837. Owned<IWUResult> r = updateWorkUnitResult(w, name, sequence);
  838. if (r)
  839. {
  840. r->setResultRaw(len, result, ResultFormatRaw);
  841. r->setResultStatus(ResultStatusCalculated);
  842. }
  843. else
  844. throw MakeStringException(TE_UnexpectedParameters, "Unexpected parameters to setResultData");
  845. }
  846. virtual void setResultReal(const char *name, unsigned sequence, double result) override
  847. {
  848. WorkunitUpdate w(&workunit->lock());
  849. Owned<IWUResult> r = updateWorkUnitResult(w, name, sequence);
  850. if (r)
  851. {
  852. r->setResultReal(result);
  853. r->setResultStatus(ResultStatusCalculated);
  854. }
  855. else
  856. throw MakeStringException(TE_UnexpectedParameters, "Unexpected parameters to setResultReal");
  857. }
  858. virtual void setResultSet(const char *name, unsigned sequence, bool isAll, size32_t len, const void *result, ISetToXmlTransformer *) override
  859. {
  860. WorkunitUpdate w(&workunit->lock());
  861. Owned<IWUResult> r = updateWorkUnitResult(w, name, sequence);
  862. if (r)
  863. {
  864. r->setResultIsAll(isAll);
  865. r->setResultRaw(len, result, ResultFormatRaw);
  866. r->setResultStatus(ResultStatusCalculated);
  867. }
  868. else
  869. throw MakeStringException(TE_UnexpectedParameters, "Unexpected parameters to setResultData");
  870. }
  871. virtual void setResultString(const char *name, unsigned sequence, int len, const char *result) override
  872. {
  873. WorkunitUpdate w(&workunit->lock());
  874. Owned<IWUResult> r = updateWorkUnitResult(w, name, sequence);
  875. if (r)
  876. {
  877. r->setResultString(result, len);
  878. r->setResultStatus(ResultStatusCalculated);
  879. }
  880. else
  881. throw MakeStringException(TE_UnexpectedParameters, "Unexpected parameters to setResultString");
  882. }
  883. virtual void setResultUnicode(const char * name, unsigned sequence, int len, UChar const * result) override
  884. {
  885. WorkunitUpdate w(&workunit->lock());
  886. Owned<IWUResult> r = updateWorkUnitResult(w, name, sequence);
  887. if (r)
  888. {
  889. r->setResultUnicode(result, len);
  890. r->setResultStatus(ResultStatusCalculated);
  891. }
  892. else
  893. throw MakeStringException(TE_UnexpectedParameters, "Unexpected parameters to setResultUnicode");
  894. }
  895. virtual void setResultVarString(const char * name, unsigned sequence, const char *result) override
  896. {
  897. WorkunitUpdate w(&workunit->lock());
  898. Owned<IWUResult> r = updateWorkUnitResult(w, name, sequence);
  899. if (r)
  900. {
  901. r->setResultString(result, strlen(result));
  902. r->setResultStatus(ResultStatusCalculated);
  903. }
  904. else
  905. throw MakeStringException(TE_UnexpectedParameters, "Unexpected parameters to setResultVarString");
  906. }
  907. virtual void setResultUInt(const char *name, unsigned sequence, unsigned __int64 result, unsigned size) override
  908. {
  909. WorkunitUpdate w(&workunit->lock());
  910. Owned<IWUResult> r = updateWorkUnitResult(w, name, sequence);
  911. if (r)
  912. {
  913. r->setResultUInt(result);
  914. r->setResultStatus(ResultStatusCalculated);
  915. }
  916. else
  917. throw MakeStringException(TE_UnexpectedParameters, "Unexpected parameters to setResultUInt");
  918. }
  919. virtual void setResultVarUnicode(const char * stepname, unsigned sequence, UChar const *val) override
  920. {
  921. setResultUnicode(stepname, sequence, rtlUnicodeStrlen(val), val);
  922. }
  923. virtual bool getResultBool(const char * stepname, unsigned sequence) override
  924. {
  925. PROTECTED_GETRESULT(stepname, sequence, "Bool", "bool",
  926. return r->getResultBool();
  927. );
  928. }
  929. virtual void getResultData(unsigned & tlen, void * & tgt, const char * stepname, unsigned sequence) override
  930. {
  931. PROTECTED_GETRESULT(stepname, sequence, "Data", "data",
  932. SCMStringBuffer result;
  933. r->getResultString(result, false);
  934. tlen = result.length();
  935. tgt = (char *)result.s.detach();
  936. );
  937. }
  938. virtual void getResultDecimal(unsigned tlen, int precision, bool isSigned, void * tgt, const char * stepname, unsigned sequence) override
  939. {
  940. PROTECTED_GETRESULT(stepname, sequence, "Decimal", "decimal",
  941. r->getResultDecimal(tgt, tlen, precision, isSigned);
  942. );
  943. }
  944. virtual void getResultRaw(unsigned & tlen, void * & tgt, const char * stepname, unsigned sequence, IXmlToRowTransformer * xmlTransformer, ICsvToRowTransformer * csvTransformer) override
  945. {
  946. tgt = NULL;
  947. PROTECTED_GETRESULT(stepname, sequence, "Raw", "raw",
  948. Variable2IDataVal result(&tlen, &tgt);
  949. Owned<IXmlToRawTransformer> rawXmlTransformer = createXmlRawTransformer(xmlTransformer);
  950. Owned<ICsvToRawTransformer> rawCsvTransformer = createCsvRawTransformer(csvTransformer);
  951. r->getResultRaw(result, rawXmlTransformer, rawCsvTransformer);
  952. );
  953. }
  954. virtual void getResultSet(bool & isAll, unsigned & tlen, void * & tgt, const char * stepname, unsigned sequence, IXmlToRowTransformer * xmlTransformer, ICsvToRowTransformer * csvTransformer) override
  955. {
  956. tgt = NULL;
  957. PROTECTED_GETRESULT(stepname, sequence, "Raw", "raw",
  958. Variable2IDataVal result(&tlen, &tgt);
  959. Owned<IXmlToRawTransformer> rawXmlTransformer = createXmlRawTransformer(xmlTransformer);
  960. Owned<ICsvToRawTransformer> rawCsvTransformer = createCsvRawTransformer(csvTransformer);
  961. isAll = r->getResultIsAll();
  962. r->getResultRaw(result, rawXmlTransformer, rawCsvTransformer);
  963. );
  964. }
  965. virtual __int64 getResultInt(const char * name, unsigned sequence) override
  966. {
  967. PROTECTED_GETRESULT(name, sequence, "Int", "integer",
  968. return r->getResultInt();
  969. );
  970. }
  971. virtual double getResultReal(const char * name, unsigned sequence) override
  972. {
  973. PROTECTED_GETRESULT(name, sequence, "Real", "real",
  974. return r->getResultReal();
  975. );
  976. }
  977. virtual void getResultString(unsigned & tlen, char * & tgt, const char * stepname, unsigned sequence) override
  978. {
  979. PROTECTED_GETRESULT(stepname, sequence, "String", "string",
  980. SCMStringBuffer result;
  981. r->getResultString(result, false);
  982. tlen = result.length();
  983. tgt = (char *)result.s.detach();
  984. );
  985. }
  986. virtual void getResultStringF(unsigned tlen, char * tgt, const char * stepname, unsigned sequence) override
  987. {
  988. PROTECTED_GETRESULT(stepname, sequence, "String", "string",
  989. SCMStringBuffer result;
  990. r->getResultString(result, false);
  991. rtlStrToStr(tlen, tgt, result.length(), result.s.str());
  992. );
  993. }
  994. virtual void getResultUnicode(unsigned & tlen, UChar * & tgt, const char * stepname, unsigned sequence) override
  995. {
  996. PROTECTED_GETRESULT(stepname, sequence, "Unicode", "unicode",
  997. MemoryBuffer result;
  998. r->getResultUnicode(MemoryBuffer2IDataVal(result));
  999. tlen = result.length()/2;
  1000. tgt = (UChar *)malloc(tlen*2);
  1001. memcpy(tgt, result.toByteArray(), tlen*2);
  1002. );
  1003. }
  1004. virtual char * getResultVarString(const char * stepname, unsigned sequence) override
  1005. {
  1006. PROTECTED_GETRESULT(stepname, sequence, "VarString", "string",
  1007. SCMStringBuffer result;
  1008. r->getResultString(result, false);
  1009. return result.s.detach();
  1010. );
  1011. }
  1012. virtual UChar * getResultVarUnicode(const char * stepname, unsigned sequence) override
  1013. {
  1014. PROTECTED_GETRESULT(stepname, sequence, "VarUnicode", "unicode",
  1015. MemoryBuffer result;
  1016. r->getResultUnicode(MemoryBuffer2IDataVal(result));
  1017. unsigned tlen = result.length()/2;
  1018. result.append((UChar)0);
  1019. return (UChar *)result.detach();
  1020. );
  1021. }
  1022. virtual unsigned getResultHash(const char * name, unsigned sequence) override
  1023. {
  1024. PROTECTED_GETRESULT(name, sequence, "Hash", "hash",
  1025. return r->getResultHash();
  1026. );
  1027. }
  1028. virtual unsigned getExternalResultHash(const char * wuid, const char * stepname, unsigned sequence) override
  1029. {
  1030. try
  1031. {
  1032. LOG(MCdebugProgress, unknownJob, "getExternalResultRaw %s", stepname);
  1033. Owned<IConstWUResult> r = getExternalResult(wuid, stepname, sequence);
  1034. return r->getResultHash();
  1035. }
  1036. catch (CATCHALL)
  1037. {
  1038. throw MakeStringException(TE_FailedToRetrieveWorkunitValue, "Failed to retrieve external data hash %s from workunit %s", stepname, wuid);
  1039. }
  1040. }
  1041. virtual void getResultRowset(size32_t & tcount, const byte * * & tgt, const char * stepname, unsigned sequence, IEngineRowAllocator * _rowAllocator, bool isGrouped, IXmlToRowTransformer * xmlTransformer, ICsvToRowTransformer * csvTransformer) override
  1042. {
  1043. tgt = NULL;
  1044. PROTECTED_GETRESULT(stepname, sequence, "Rowset", "rowset",
  1045. MemoryBuffer datasetBuffer;
  1046. MemoryBuffer2IDataVal result(datasetBuffer);
  1047. Owned<IXmlToRawTransformer> rawXmlTransformer = createXmlRawTransformer(xmlTransformer);
  1048. Owned<ICsvToRawTransformer> rawCsvTransformer = createCsvRawTransformer(csvTransformer);
  1049. r->getResultRaw(result, rawXmlTransformer, rawCsvTransformer);
  1050. Owned<IOutputRowDeserializer> deserializer = _rowAllocator->createDiskDeserializer(this);
  1051. rtlDataset2RowsetX(tcount, tgt, _rowAllocator, deserializer, datasetBuffer.length(), datasetBuffer.toByteArray(), isGrouped);
  1052. );
  1053. }
  1054. virtual void getResultDictionary(size32_t & tcount, const byte * * & tgt, IEngineRowAllocator * _rowAllocator, const char * stepname, unsigned sequence, IXmlToRowTransformer * xmlTransformer, ICsvToRowTransformer * csvTransformer, IHThorHashLookupInfo * hasher) override
  1055. {
  1056. tcount = 0;
  1057. tgt = NULL;
  1058. PROTECTED_GETRESULT(stepname, sequence, "Dictionary", "dictionary",
  1059. MemoryBuffer datasetBuffer;
  1060. MemoryBuffer2IDataVal result(datasetBuffer);
  1061. Owned<IXmlToRawTransformer> rawXmlTransformer = createXmlRawTransformer(xmlTransformer);
  1062. Owned<ICsvToRawTransformer> rawCsvTransformer = createCsvRawTransformer(csvTransformer);
  1063. r->getResultRaw(result, rawXmlTransformer, rawCsvTransformer);
  1064. Owned<IOutputRowDeserializer> deserializer = _rowAllocator->createDiskDeserializer(this);
  1065. rtlDeserializeDictionary(tcount, tgt, _rowAllocator, deserializer, datasetBuffer.length(), datasetBuffer.toByteArray());
  1066. );
  1067. }
  1068. virtual void getExternalResultRaw(unsigned & tlen, void * & tgt, const char * wuid, const char * stepname, unsigned sequence, IXmlToRowTransformer * xmlTransformer, ICsvToRowTransformer * csvTransformer) override
  1069. {
  1070. tgt = NULL;
  1071. try
  1072. {
  1073. LOG(MCdebugProgress, unknownJob, "getExternalResultRaw %s", stepname);
  1074. Variable2IDataVal result(&tlen, &tgt);
  1075. Owned<IConstWUResult> r = getExternalResult(wuid, stepname, sequence);
  1076. Owned<IXmlToRawTransformer> rawXmlTransformer = createXmlRawTransformer(xmlTransformer);
  1077. Owned<ICsvToRawTransformer> rawCsvTransformer = createCsvRawTransformer(csvTransformer);
  1078. r->getResultRaw(result, rawXmlTransformer, rawCsvTransformer);
  1079. }
  1080. catch (CATCHALL)
  1081. {
  1082. throw MakeStringException(TE_FailedToRetrieveWorkunitValue, "Failed to retrieve external data value %s from workunit %s", stepname, wuid);
  1083. }
  1084. }
  1085. virtual void addWuException(const char * text, unsigned code, unsigned severity, const char * source) override
  1086. {
  1087. addWuExceptionEx(text, code, severity, MSGAUD_programmer, source);
  1088. }
  1089. virtual void addWuAssertFailure(unsigned code, const char * text, const char * filename, unsigned lineno, unsigned column, bool isAbort) override
  1090. {
  1091. DBGLOG("%s", text);
  1092. try
  1093. {
  1094. Owned<IWorkUnit> w = updateWorkUnit();
  1095. unsigned activity = 0;
  1096. addExceptionToWorkunit(w, SeverityError, "user", code, text, filename, lineno, column, activity);
  1097. }
  1098. catch (IException *E)
  1099. {
  1100. StringBuffer m;
  1101. E->errorMessage(m);
  1102. IERRLOG("Unable to record exception in workunit: %s", m.str());
  1103. E->Release();
  1104. }
  1105. catch (...)
  1106. {
  1107. IERRLOG("Unable to record exception in workunit: unknown exception");
  1108. }
  1109. if (isAbort)
  1110. throw makeStringException(MSGAUD_user, code, text);
  1111. }
  1112. virtual unsigned __int64 getFileOffset(const char *logicalName) override { assertex(false); return 0; }
  1113. virtual unsigned getNodes() override { return jobChannel.queryJob().querySlaves(); }
  1114. virtual unsigned getNodeNum() override { throw MakeThorException(0, "Unsupported. getNodeNum() called in master"); return (unsigned)-1; }
  1115. virtual char *getFilePart(const char *logicalName, bool create=false) override { assertex(false); return NULL; }
  1116. virtual unsigned __int64 getDatasetHash(const char * name, unsigned __int64 hash) override
  1117. {
  1118. unsigned checkSum = 0;
  1119. Owned<IDistributedFile> iDfsFile = queryThorFileManager().lookup(jobChannel.queryJob(), name, false, true, false, defaultPrivilegedUser); // NB: do not update accessed
  1120. if (iDfsFile.get())
  1121. {
  1122. if (iDfsFile->getFileCheckSum(checkSum))
  1123. hash ^= checkSum;
  1124. else
  1125. {
  1126. StringBuffer modifiedStr;
  1127. if (iDfsFile->queryAttributes().getProp("@modified", modifiedStr))
  1128. hash = rtlHash64Data(modifiedStr.length(), modifiedStr.str(), hash);
  1129. // JCS->GH - what's the best thing to do here, if [for some reason] neither are available..
  1130. }
  1131. }
  1132. return hash;
  1133. }
  1134. virtual IDistributedFileTransaction *querySuperFileTransaction() override
  1135. {
  1136. if (!superfiletransaction.get())
  1137. superfiletransaction.setown(createDistributedFileTransaction(userDesc, this));
  1138. return superfiletransaction.get();
  1139. }
  1140. virtual char *getJobName() override
  1141. {
  1142. return strdup(workunit->queryJobName());
  1143. }
  1144. virtual char *getClusterName() override
  1145. {
  1146. return strdup(workunit->queryClusterName());
  1147. }
  1148. virtual char *getGroupName() override
  1149. {
  1150. StringBuffer out;
  1151. if (globals)
  1152. globals->getProp("@nodeGroup",out);
  1153. return out.detach();
  1154. }
  1155. // ICodeContextExt impl.
  1156. virtual IConstWUResult *getExternalResult(const char * wuid, const char *name, unsigned sequence) override
  1157. {
  1158. Owned<IWorkUnitFactory> factory = getWorkUnitFactory();
  1159. Owned<IConstWorkUnit> externalWU = factory->openWorkUnit(wuid);
  1160. if (!externalWU)
  1161. throw MakeStringException(TE_FailedToRetrieveWorkunitValue, "workunit %s not found, retrieving value %s", wuid, name);
  1162. externalWU->remoteCheckAccess(userDesc, false);
  1163. return getWorkUnitResult(externalWU, name, sequence);
  1164. }
  1165. virtual IConstWUResult *getResultForGet(const char *name, unsigned sequence) override
  1166. {
  1167. Owned<IConstWUResult> r = getResult(name, sequence);
  1168. if (!r || (r->getResultStatus() == ResultStatusUndefined))
  1169. {
  1170. StringBuffer s;
  1171. throw MakeStringException(TE_FailedToRetrieveWorkunitValue, "value %s in workunit is undefined", getResultText(s,name,sequence));
  1172. }
  1173. return r.getClear();
  1174. }
  1175. virtual IWorkUnit *updateWorkUnit() const override
  1176. {
  1177. return &workunit->lock();
  1178. }
  1179. virtual void addWuExceptionEx(const char * text, unsigned code, unsigned severity, unsigned audience, const char * source) override
  1180. {
  1181. LOG(mapToLogMsgCategory((ErrorSeverity)severity, (MessageAudience)audience), "%s", text);
  1182. try
  1183. {
  1184. Owned<IWorkUnit> w = updateWorkUnit();
  1185. Owned<IWUException> we = w->createException();
  1186. we->setSeverity((ErrorSeverity)severity);
  1187. we->setExceptionMessage(text);
  1188. we->setExceptionSource(source);
  1189. if (code)
  1190. we->setExceptionCode(code);
  1191. }
  1192. catch (IException *E)
  1193. {
  1194. StringBuffer m;
  1195. E->errorMessage(m);
  1196. IERRLOG("Unable to record exception in workunit: %s", m.str());
  1197. E->Release();
  1198. }
  1199. catch (...)
  1200. {
  1201. IERRLOG("Unable to record exception in workunit: unknown exception");
  1202. }
  1203. }
  1204. };
  1205. class CThorCodeContextMasterSharedMem : public CThorCodeContextMaster
  1206. {
  1207. IThorAllocator *sharedAllocator;
  1208. public:
  1209. CThorCodeContextMasterSharedMem(CJobChannel &jobChannel, IThorAllocator *_sharedAllocator, IConstWorkUnit &_workunit, ILoadedDllEntry &querySo, IUserDescriptor &userDesc)
  1210. : CThorCodeContextMaster(jobChannel, _workunit, querySo, userDesc)
  1211. {
  1212. sharedAllocator = _sharedAllocator;
  1213. }
  1214. virtual IEngineRowAllocator *getRowAllocator(IOutputMetaData * meta, unsigned activityId) const
  1215. {
  1216. return sharedAllocator->getRowAllocator(meta, activityId);
  1217. }
  1218. };
  1219. /////////////
  1220. //
  1221. // CJobMaster
  1222. //
  1223. void loadPlugin(SafePluginMap *pluginMap, const char *_path, const char *name)
  1224. {
  1225. StringBuffer path(_path);
  1226. path.append(name);
  1227. pluginMap->addPlugin(path.str(), name);
  1228. }
  1229. CJobMaster::CJobMaster(IConstWorkUnit &_workunit, const char *graphName, ILoadedDllEntry *querySo, bool _sendSo, const SocketEndpoint &_agentEp)
  1230. : CJobBase(querySo, graphName), workunit(&_workunit), sendSo(_sendSo), agentEp(_agentEp)
  1231. {
  1232. SCMStringBuffer _token, _scope;
  1233. workunit->getScope(_scope);
  1234. workunit->getWorkunitDistributedAccessToken(_token);
  1235. wuid.set(workunit->queryWuid());
  1236. user.set(workunit->queryUser());
  1237. token.append(_token.str());
  1238. scope.append(_scope.str());
  1239. globalMemoryMB = globals->getPropInt("@masterMemorySize", globals->getPropInt("@globalMemorySize")); // in MB
  1240. numChannels = 1;
  1241. init();
  1242. if (workunit->hasDebugValue("GlobalId"))
  1243. {
  1244. SCMStringBuffer txId;
  1245. workunit->getDebugValue("GlobalId", txId);
  1246. if (txId.length())
  1247. {
  1248. SocketEndpoint thorEp;
  1249. thorEp.setLocalHost(getMachinePortBase());
  1250. logctx->setGlobalId(txId.str(), thorEp, 0);
  1251. SCMStringBuffer callerId;
  1252. workunit->getDebugValue("CallerId", callerId);
  1253. if (callerId.length())
  1254. logctx->setCallerId(callerId.str());
  1255. VStringBuffer msg("GlobalId: %s", txId.str());
  1256. workunit->getDebugValue("CallerId", txId);
  1257. if (txId.length())
  1258. msg.append(", CallerId: ").append(txId.str());
  1259. txId.set(logctx->queryLocalId());
  1260. if (txId.length())
  1261. msg.append(", LocalId: ").append(txId.str());
  1262. logctx->CTXLOG("%s", msg.str());
  1263. }
  1264. }
  1265. resumed = WUActionResume == workunit->getAction();
  1266. fatalHandler.setown(new CFatalHandler(globals->getPropInt("@fatal_timeout", FATAL_TIMEOUT)));
  1267. querySent = spillsSaved = false;
  1268. nodeDiskUsageCached = false;
  1269. StringBuffer pluginsDir;
  1270. globals->getProp("@pluginsPath", pluginsDir);
  1271. if (pluginsDir.length())
  1272. addPathSepChar(pluginsDir);
  1273. Owned<IConstWUPluginIterator> pluginIter = &workunit->getPlugins();
  1274. ForEach(*pluginIter)
  1275. {
  1276. IConstWUPlugin &plugin = pluginIter->query();
  1277. SCMStringBuffer name;
  1278. plugin.getPluginName(name);
  1279. loadPlugin(pluginMap, pluginsDir.str(), name.str());
  1280. }
  1281. sharedAllocator.setown(::createThorAllocator(globalMemoryMB, 0, 1, memorySpillAtPercentage, *logctx, crcChecking, usePackedAllocator));
  1282. Owned<IMPServer> mpServer = getMPServer();
  1283. addChannel(mpServer);
  1284. slavemptag = allocateMPTag();
  1285. slaveMsgHandler.setown(new CSlaveMessageHandler(*this, slavemptag));
  1286. tmpHandler.setown(createTempHandler(true));
  1287. xgmml.set(graphXGMML);
  1288. }
  1289. void CJobMaster::endJob()
  1290. {
  1291. if (jobEnded)
  1292. return;
  1293. slaveMsgHandler.clear();
  1294. freeMPTag(slavemptag);
  1295. tmpHandler.clear();
  1296. PARENT::endJob();
  1297. }
  1298. void CJobMaster::addChannel(IMPServer *mpServer)
  1299. {
  1300. jobChannels.append(*new CJobMasterChannel(*this, mpServer, jobChannels.ordinality()));
  1301. }
  1302. static IException *createBCastException(unsigned slave, const char *errorMsg)
  1303. {
  1304. // think this should always be fatal, could check link down here, or in general and flag as _shutdown.
  1305. StringBuffer msg("General failure communicating to slave");
  1306. if (slave)
  1307. msg.append("(").append(slave).append(") ");
  1308. else
  1309. msg.append("s ");
  1310. Owned<IThorException> e = MakeThorException(0, "%s", msg.append(" [").append(errorMsg).append("]").str());
  1311. e->setAction(tea_shutdown);
  1312. return e.getClear();
  1313. }
  1314. mptag_t CJobMaster::allocateMPTag()
  1315. {
  1316. mptag_t tag = allocateClusterMPTag();
  1317. queryJobChannel(0).queryJobComm().flush(tag);
  1318. PROGLOG("allocateMPTag: tag = %d", (int)tag);
  1319. return tag;
  1320. }
  1321. void CJobMaster::freeMPTag(mptag_t tag)
  1322. {
  1323. if (TAG_NULL != tag)
  1324. {
  1325. freeClusterMPTag(tag);
  1326. PROGLOG("freeMPTag: tag = %d", (int)tag);
  1327. queryJobChannel(0).queryJobComm().flush(tag);
  1328. }
  1329. }
  1330. void CJobMaster::broadcast(ICommunicator &comm, CMessageBuffer &msg, mptag_t mptag, unsigned timeout, const char *errorMsg, CReplyCancelHandler *msgHandler, bool sendOnly)
  1331. {
  1332. unsigned groupSizeExcludingMaster = comm.queryGroup().ordinality() - 1;
  1333. mptag_t replyTag = TAG_NULL;
  1334. if (!sendOnly)
  1335. {
  1336. replyTag = queryJobChannel(0).queryMPServer().createReplyTag();
  1337. msg.setReplyTag(replyTag);
  1338. }
  1339. if (globals->getPropBool("@broadcastSendAsync", true)) // only here in case of problems/debugging.
  1340. {
  1341. class CSendAsyncfor : public CAsyncFor
  1342. {
  1343. CMessageBuffer &msg;
  1344. mptag_t mptag;
  1345. unsigned timeout;
  1346. StringAttr errorMsg;
  1347. ICommunicator &comm;
  1348. public:
  1349. CSendAsyncfor(ICommunicator &_comm, CMessageBuffer &_msg, mptag_t _mptag, unsigned _timeout, const char *_errorMsg)
  1350. : comm(_comm), msg(_msg), mptag(_mptag), timeout(_timeout), errorMsg(_errorMsg)
  1351. {
  1352. }
  1353. void Do(unsigned i)
  1354. {
  1355. if (!comm.send(msg, i+1, mptag, timeout))
  1356. throw createBCastException(i+1, errorMsg);
  1357. }
  1358. } afor(comm, msg, mptag, timeout, errorMsg);
  1359. try
  1360. {
  1361. afor.For(groupSizeExcludingMaster, groupSizeExcludingMaster);
  1362. }
  1363. catch (IException *e)
  1364. {
  1365. EXCLOG(e, "broadcastSendAsync");
  1366. abort(e);
  1367. throw;
  1368. }
  1369. }
  1370. else if (!comm.send(msg, RANK_ALL_OTHER, mptag, timeout))
  1371. {
  1372. Owned<IException> e = createBCastException(0, errorMsg);
  1373. EXCLOG(e, NULL);
  1374. abort(e);
  1375. throw e.getClear();
  1376. }
  1377. if (sendOnly) return;
  1378. unsigned respondents = 0;
  1379. Owned<IBitSet> bitSet = createThreadSafeBitSet();
  1380. for (;;)
  1381. {
  1382. rank_t sender;
  1383. CMessageBuffer msg;
  1384. bool r = msgHandler ? msgHandler->recv(comm, msg, RANK_ALL, replyTag, &sender, LONGTIMEOUT)
  1385. : comm.recv(msg, RANK_ALL, replyTag, &sender, LONGTIMEOUT);
  1386. if (!r)
  1387. {
  1388. StringBuffer tmpStr;
  1389. if (errorMsg)
  1390. tmpStr.append(": ").append(errorMsg).append(" - ");
  1391. tmpStr.append("Timeout receiving from slaves - no reply from: [");
  1392. unsigned s = bitSet->scan(0, false);
  1393. assertex(s<querySlaves()); // must be at least one
  1394. tmpStr.append(s+1);
  1395. for (;;)
  1396. {
  1397. s = bitSet->scan(s+1, false);
  1398. if (s>=querySlaves())
  1399. break;
  1400. tmpStr.append(",").append(s+1);
  1401. }
  1402. tmpStr.append("]");
  1403. Owned<IException> e = MakeThorFatal(NULL, 0, " %s", tmpStr.str());
  1404. EXCLOG(e, NULL);
  1405. throw e.getClear();
  1406. }
  1407. bool error;
  1408. msg.read(error);
  1409. if (error)
  1410. {
  1411. Owned<IThorException> e = deserializeThorException(msg);
  1412. e->setSlave(sender);
  1413. throw e.getClear();
  1414. }
  1415. ++respondents;
  1416. bitSet->set((unsigned)sender-1);
  1417. if (respondents == groupSizeExcludingMaster)
  1418. break;
  1419. }
  1420. }
  1421. void CJobMaster::initNodeDUCache()
  1422. {
  1423. if (!nodeDiskUsageCached)
  1424. {
  1425. nodeDiskUsageCached = true;
  1426. Owned<IPropertyTreeIterator> fileIter = &workunit->getFileIterator();
  1427. ForEach (*fileIter)
  1428. {
  1429. Owned<IDistributedFile> f = queryDistributedFileDirectory().lookup(fileIter->query().queryProp("@name"), userDesc, false, false, false, nullptr, defaultPrivilegedUser);
  1430. if (f)
  1431. {
  1432. unsigned n = f->numParts();
  1433. for (unsigned i=0;i<n;i++)
  1434. {
  1435. Owned<IDistributedFilePart> part = f->getPart(i);
  1436. offset_t sz = part->getFileSize(false, false);
  1437. if (i>=nodeDiskUsage.ordinality())
  1438. nodeDiskUsage.append(sz);
  1439. else
  1440. {
  1441. sz += nodeDiskUsage.item(i);
  1442. nodeDiskUsage.add(sz, i);
  1443. }
  1444. }
  1445. }
  1446. }
  1447. }
  1448. }
  1449. IPropertyTree *CJobMaster::prepareWorkUnitInfo()
  1450. {
  1451. Owned<IPropertyTree> workUnitInfo = createPTree("workUnitInfo");
  1452. workUnitInfo->setProp("wuid", wuid);
  1453. workUnitInfo->setProp("user", user);
  1454. workUnitInfo->setProp("token", token);
  1455. workUnitInfo->setProp("scope", scope);
  1456. Owned<IConstWUPluginIterator> pluginIter = &queryWorkUnit().getPlugins();
  1457. IPropertyTree *plugins = NULL;
  1458. ForEach(*pluginIter)
  1459. {
  1460. IConstWUPlugin &thisplugin = pluginIter->query();
  1461. if (!plugins)
  1462. plugins = workUnitInfo->addPropTree("plugins", createPTree());
  1463. SCMStringBuffer name;
  1464. thisplugin.getPluginName(name);
  1465. IPropertyTree *plugin = plugins->addPropTree("plugin", createPTree());
  1466. plugin->setProp("@name", name.str());
  1467. }
  1468. IPropertyTree *debug = workUnitInfo->addPropTree("Debug", createPTree(ipt_caseInsensitive));
  1469. SCMStringBuffer debugStr, valueStr;
  1470. Owned<IStringIterator> debugIter = &queryWorkUnit().getDebugValues();
  1471. ForEach (*debugIter)
  1472. {
  1473. debugIter->str(debugStr);
  1474. queryWorkUnit().getDebugValue(debugStr.str(), valueStr);
  1475. debug->setProp(debugStr.str(), valueStr.str());
  1476. }
  1477. return workUnitInfo.getClear();
  1478. }
  1479. void CJobMaster::sendQuery()
  1480. {
  1481. CriticalBlock b(sendQueryCrit);
  1482. if (querySent) return;
  1483. CMessageBuffer tmp;
  1484. tmp.append(slavemptag);
  1485. tmp.append(queryWuid());
  1486. tmp.append(graphName);
  1487. const char *soName = queryDllEntry().queryName();
  1488. PROGLOG("Query dll: %s", soName);
  1489. tmp.append(soName);
  1490. tmp.append(sendSo);
  1491. if (sendSo)
  1492. {
  1493. CTimeMon atimer;
  1494. OwnedIFile iFile = createIFile(soName);
  1495. OwnedIFileIO iFileIO = iFile->open(IFOread);
  1496. size32_t sz = (size32_t)iFileIO->size();
  1497. tmp.append(sz);
  1498. read(iFileIO, 0, sz, tmp);
  1499. PROGLOG("Loading query for serialization to slaves took %d ms", atimer.elapsed());
  1500. }
  1501. Owned<IPropertyTree> deps = createPTree(queryXGMML()->queryName());
  1502. Owned<IPropertyTreeIterator> edgeIter = queryXGMML()->getElements("edge"); // JCSMORE trim to those actually needed
  1503. ForEach (*edgeIter)
  1504. {
  1505. IPropertyTree &edge = edgeIter->query();
  1506. deps->addPropTree("edge", LINK(&edge));
  1507. }
  1508. Owned<IPropertyTree> workUnitInfo = prepareWorkUnitInfo();
  1509. workUnitInfo->serialize(tmp);
  1510. deps->serialize(tmp);
  1511. CMessageBuffer msg;
  1512. msg.append(QueryInit);
  1513. compressToBuffer(msg, tmp.length(), tmp.toByteArray());
  1514. CTimeMon queryToSlavesTimer;
  1515. broadcast(queryNodeComm(), msg, masterSlaveMpTag, LONGTIMEOUT, "sendQuery");
  1516. PROGLOG("Serialization of query init info (%d bytes) to slaves took %d ms", msg.length(), queryToSlavesTimer.elapsed());
  1517. queryJobManager().addCachedSo(soName);
  1518. querySent = true;
  1519. }
  1520. void CJobMaster::jobDone()
  1521. {
  1522. if (!querySent) return;
  1523. CMessageBuffer msg;
  1524. msg.append(QueryDone);
  1525. msg.append(queryKey());
  1526. broadcast(queryNodeComm(), msg, masterSlaveMpTag, LONGTIMEOUT, "jobDone");
  1527. }
  1528. void CJobMaster::saveSpills()
  1529. {
  1530. CriticalBlock b(spillCrit);
  1531. if (spillsSaved)
  1532. return;
  1533. spillsSaved = true;
  1534. PROGLOG("Paused, saving spills..");
  1535. assertex(!queryUseCheckpoints()); // JCSMORE - checkpoints probably need revisiting
  1536. unsigned numSavedSpills = 0;
  1537. // stash away spills ready for resume, make them owned by workunit in event of abort/delete
  1538. IGraphTempHandler *tempHandler = queryTempHandler(false);
  1539. if (tempHandler)
  1540. {
  1541. Owned<IFileUsageIterator> iter = tempHandler->getIterator();
  1542. ForEach(*iter)
  1543. {
  1544. CFileUsageEntry &entry = iter->query();
  1545. StringAttr tmpName = entry.queryName();
  1546. if (WUGraphComplete == workunit->queryNodeState(queryGraphName(), entry.queryGraphId()))
  1547. {
  1548. IArrayOf<IGroup> groups;
  1549. StringArray clusters;
  1550. fillClusterArray(*this, tmpName, clusters, groups);
  1551. Owned<IFileDescriptor> fileDesc = queryThorFileManager().create(*this, tmpName, clusters, groups, true, TDXtemporary|TDWnoreplicate);
  1552. fileDesc->queryProperties().setPropBool("@pausefile", true); // JCSMORE - mark to keep, may be able to distinguish via other means
  1553. fileDesc->queryProperties().setProp("@kind", "flat");
  1554. IPropertyTree &props = fileDesc->queryProperties();
  1555. props.setPropBool("@owned", true);
  1556. bool blockCompressed=true; // JCSMORE, should come from helper really
  1557. if (blockCompressed)
  1558. props.setPropBool("@blockCompressed", true);
  1559. Owned<IDistributedFile> file = queryDistributedFileDirectory().createNew(fileDesc);
  1560. // NB: This is renaming/moving from temp path
  1561. StringBuffer newName;
  1562. queryThorFileManager().addScope(*this, tmpName, newName, true, true);
  1563. verifyex(file->renamePhysicalPartFiles(newName.str(), NULL, NULL, queryBaseDirectory(grp_unknown)));
  1564. file->attach(newName,userDesc);
  1565. Owned<IWorkUnit> wu = &queryWorkUnit().lock();
  1566. wu->addFile(newName, &clusters, entry.queryUsage(), entry.queryKind(), queryGraphName());
  1567. ++numSavedSpills;
  1568. }
  1569. }
  1570. }
  1571. PROGLOG("Paused, %d spill(s) saved.", numSavedSpills);
  1572. }
  1573. bool CJobMaster::go()
  1574. {
  1575. class CWorkunitPauseHandler : public CInterface, implements IWorkUnitSubscriber
  1576. {
  1577. CJobMaster &job;
  1578. IConstWorkUnit &wu;
  1579. Owned<IWorkUnitWatcher> watcher;
  1580. CriticalSection crit;
  1581. public:
  1582. IMPLEMENT_IINTERFACE;
  1583. CWorkunitPauseHandler(CJobMaster &_job, IConstWorkUnit &_wu) : job(_job), wu(_wu)
  1584. {
  1585. Owned<IWorkUnitFactory> factory = getWorkUnitFactory();
  1586. watcher.setown(factory->getWatcher(this, (WUSubscribeOptions) (SubscribeOptionAction | SubscribeOptionAbort), wu.queryWuid()));
  1587. }
  1588. ~CWorkunitPauseHandler() { stop(); }
  1589. void stop()
  1590. {
  1591. Owned<IWorkUnitWatcher> _watcher;
  1592. {
  1593. CriticalBlock b(crit);
  1594. if (!watcher)
  1595. return;
  1596. _watcher.setown(watcher.getClear());
  1597. }
  1598. _watcher->unsubscribe();
  1599. }
  1600. virtual void notify(WUSubscribeOptions flags, unsigned valueLen, const void *valueData) override
  1601. {
  1602. CriticalBlock b(crit);
  1603. if (!watcher)
  1604. return;
  1605. if (flags & SubscribeOptionAbort)
  1606. {
  1607. Owned<IWorkUnitFactory> factory = getWorkUnitFactory();
  1608. if (factory->isAborting(wu.queryWuid()))
  1609. {
  1610. LOG(MCwarning, thorJob, "ABORT detected from user");
  1611. unsigned code = TE_WorkUnitAborting; // default
  1612. if (job.getOptBool("dumpInfoOnUserAbort", false))
  1613. code = TE_WorkUnitAbortingDumpInfo;
  1614. else if ((1 == valueLen) && ('2' == *((const char *)valueData)))
  1615. {
  1616. /* NB: Standard abort mechanism will trigger abort subscriber with "1"
  1617. * If "2" is signalled, it means - Abort with dump info.
  1618. */
  1619. code = TE_WorkUnitAbortingDumpInfo;
  1620. }
  1621. Owned <IException> e = MakeThorException(code, "User signalled abort");
  1622. job.fireException(e);
  1623. }
  1624. }
  1625. if (flags & SubscribeOptionAction)
  1626. {
  1627. job.markWuDirty();
  1628. bool abort = false;
  1629. bool pause = false;
  1630. wu.forceReload();
  1631. WUAction action = wu.getAction();
  1632. if (action==WUActionPause)
  1633. {
  1634. // pause after current subgraph
  1635. pause = true;
  1636. }
  1637. else if (action==WUActionPauseNow)
  1638. {
  1639. // abort current subgraph
  1640. abort = true;
  1641. pause = true;
  1642. }
  1643. if (pause)
  1644. {
  1645. PROGLOG("Pausing job%s", abort?" [now]":"");
  1646. job.pause(abort);
  1647. }
  1648. }
  1649. }
  1650. } workunitPauseHandler(*this, *workunit);
  1651. class CQueryTimeoutHandler : public CTimeoutTrigger
  1652. {
  1653. CJobMaster &job;
  1654. public:
  1655. CQueryTimeoutHandler(CJobMaster &_job, unsigned timeout) : CTimeoutTrigger(timeout, "QUERY"), job(_job)
  1656. {
  1657. inform(MakeThorException(TE_QueryTimeoutError, "Query took greater than %d seconds", timeout));
  1658. }
  1659. virtual bool action()
  1660. {
  1661. job.fireException(exception);
  1662. return true;
  1663. }
  1664. private:
  1665. graph_id graphId;
  1666. };
  1667. Owned<CTimeoutTrigger> qtHandler;
  1668. int guillotineTimeout = workunit->getDebugValueInt("maxRunTime", 0);
  1669. if (guillotineTimeout > 0)
  1670. qtHandler.setown(new CQueryTimeoutHandler(*this, guillotineTimeout));
  1671. else if (guillotineTimeout < 0)
  1672. {
  1673. Owned<IException> e = MakeStringException(0, "Ignoring negative maxRunTime: %d", guillotineTimeout);
  1674. reportExceptionToWorkunit(*workunit, e);
  1675. }
  1676. if (WUActionPause == workunit->getAction() || WUActionPauseNow == workunit->getAction())
  1677. throw MakeStringException(0, "Job paused at start, exiting");
  1678. bool allDone = true;
  1679. unsigned concurrentSubGraphs = (unsigned)getWorkUnitValueInt("concurrentSubGraphs", globals->getPropInt("@concurrentSubGraphs", 1));
  1680. try
  1681. {
  1682. startJob();
  1683. workunit->setGraphState(queryGraphName(), getWfid(), WUGraphRunning);
  1684. Owned<IThorGraphIterator> iter = queryJobChannel(0).getSubGraphs();
  1685. CICopyArrayOf<CMasterGraph> toRun;
  1686. ForEach(*iter)
  1687. {
  1688. CMasterGraph &graph = (CMasterGraph &)iter->query();
  1689. if ((queryResumed() || queryUseCheckpoints()) && WUGraphComplete == workunit->queryNodeState(queryGraphName(), graph.queryGraphId()))
  1690. graph.setCompleteEx();
  1691. else
  1692. toRun.append(graph);
  1693. }
  1694. ForEachItemInRev(g, toRun)
  1695. {
  1696. if (aborted) break;
  1697. CMasterGraph &graph = toRun.item(g);
  1698. if (graph.isSink())
  1699. graph.execute(0, NULL, true, concurrentSubGraphs>1);
  1700. if (queryPausing()) break;
  1701. }
  1702. queryJobChannel(0).wait();
  1703. workunitPauseHandler.stop();
  1704. ForEachItemIn(tr, toRun)
  1705. {
  1706. CMasterGraph &graph = toRun.item(tr);
  1707. if (!graph.isComplete())
  1708. {
  1709. allDone = false;
  1710. break;
  1711. }
  1712. }
  1713. }
  1714. catch (IException *e) { fireException(e); e->Release(); }
  1715. catch (CATCHALL) { Owned<IException> e = MakeThorException(0, "Unknown exception running sub graphs"); fireException(e); }
  1716. workunit->setGraphState(queryGraphName(), getWfid(), aborted?WUGraphFailed:(allDone?WUGraphComplete:(pausing?WUGraphPaused:WUGraphComplete)));
  1717. if (queryPausing())
  1718. saveSpills();
  1719. Owned<IException> jobDoneException;
  1720. try { jobDone(); }
  1721. catch (IException *e)
  1722. {
  1723. EXCLOG(e, NULL);
  1724. jobDoneException.setown(e);
  1725. }
  1726. fatalHandler->clear();
  1727. queryTempHandler()->clearTemps();
  1728. slaveMsgHandler->stop();
  1729. if (jobDoneException.get())
  1730. throw LINK(jobDoneException);
  1731. return allDone;
  1732. }
  1733. void CJobMaster::pause(bool doAbort)
  1734. {
  1735. pausing = true;
  1736. if (doAbort)
  1737. {
  1738. // reply will trigger DAMP_THOR_REPLY_PAUSED to agent, unless all graphs are already complete.
  1739. queryJobManager().replyException(*this, NULL);
  1740. // abort current graph asynchronously.
  1741. // After spill files have been saved, trigger timeout handler in case abort doesn't succeed.
  1742. Owned<IException> e = MakeThorException(0, "Unable to recover from pausenow");
  1743. class CAbortThread : implements IThreaded
  1744. {
  1745. CJobMaster &owner;
  1746. CThreaded threaded;
  1747. Linked<IException> exception;
  1748. public:
  1749. CAbortThread(CJobMaster &_owner, IException *_exception) : owner(_owner), exception(_exception), threaded("SaveSpillThread", this)
  1750. {
  1751. threaded.start();
  1752. }
  1753. ~CAbortThread()
  1754. {
  1755. threaded.join();
  1756. }
  1757. // IThreaded
  1758. virtual void threadmain() override
  1759. {
  1760. owner.abort(exception);
  1761. }
  1762. } abortThread(*this, e);
  1763. saveSpills();
  1764. fatalHandler->inform(e.getClear());
  1765. }
  1766. }
  1767. __int64 CJobMaster::queryNodeDiskUsage(unsigned node)
  1768. {
  1769. initNodeDUCache();
  1770. if (!nodeDiskUsage.isItem(node)) return 0;
  1771. return nodeDiskUsage.item(node);
  1772. }
  1773. void CJobMaster::setNodeDiskUsage(unsigned node, __int64 sz)
  1774. {
  1775. initNodeDUCache();
  1776. while (nodeDiskUsage.ordinality() <= node)
  1777. nodeDiskUsage.append(0);
  1778. nodeDiskUsage.replace(sz, node);
  1779. }
  1780. __int64 CJobMaster::addNodeDiskUsage(unsigned node, __int64 sz)
  1781. {
  1782. sz += queryNodeDiskUsage(node);
  1783. setNodeDiskUsage(node, sz);
  1784. return sz;
  1785. }
  1786. bool CJobMaster::queryCreatedFile(const char *file)
  1787. {
  1788. StringBuffer scopedName;
  1789. queryThorFileManager().addScope(*this, file, scopedName, false);
  1790. return (NotFound != createdFiles.find(scopedName.str()));
  1791. }
  1792. void CJobMaster::addCreatedFile(const char *file)
  1793. {
  1794. StringBuffer scopedName;
  1795. queryThorFileManager().addScope(*this, file, scopedName, false);
  1796. createdFiles.append(scopedName.str());
  1797. }
  1798. __int64 CJobMaster::getWorkUnitValueInt(const char *prop, __int64 defVal) const
  1799. {
  1800. return queryWorkUnit().getDebugValueInt64(prop, defVal);
  1801. }
  1802. bool CJobMaster::getWorkUnitValueBool(const char *prop, bool defVal) const
  1803. {
  1804. return queryWorkUnit().getDebugValueBool(prop, defVal);
  1805. }
  1806. StringBuffer &CJobMaster::getWorkUnitValue(const char *prop, StringBuffer &str) const
  1807. {
  1808. SCMStringBuffer scmStr;
  1809. queryWorkUnit().getDebugValue(prop, scmStr);
  1810. return str.append(scmStr.str());
  1811. }
  1812. IGraphTempHandler *CJobMaster::createTempHandler(bool errorOnMissing)
  1813. {
  1814. return new CMasterGraphTempHandler(*this, errorOnMissing);
  1815. }
  1816. bool CJobMaster::fireException(IException *e)
  1817. {
  1818. IThorException *te = QUERYINTERFACE(e, IThorException);
  1819. ThorExceptionAction action;
  1820. if (!te) action = tea_null;
  1821. else action = te->queryAction();
  1822. if (QUERYINTERFACE(e, IMP_Exception) && MPERR_link_closed==e->errorCode())
  1823. action = tea_shutdown;
  1824. else if (QUERYINTERFACE(e, ISEH_Exception))
  1825. action = tea_shutdown;
  1826. CriticalBlock b(exceptCrit);
  1827. switch (action)
  1828. {
  1829. case tea_warning:
  1830. {
  1831. LOG(MCwarning, thorJob, e);
  1832. ErrorSeverity mappedSeverity = workunit->getWarningSeverity(e->errorCode(), SeverityWarning);
  1833. if (mappedSeverity != SeverityIgnore)
  1834. reportExceptionToWorkunit(*workunit, e);
  1835. break;
  1836. }
  1837. default:
  1838. {
  1839. LOG(MCerror, thorJob, e);
  1840. queryJobManager().replyException(*this, e);
  1841. fatalHandler->inform(LINK(e));
  1842. try { abort(e); }
  1843. catch (IException *e)
  1844. {
  1845. Owned<IThorException> te = ThorWrapException(e, "Error aborting job, will cause thor restart");
  1846. e->Release();
  1847. reportExceptionToWorkunit(*workunit, te);
  1848. action = tea_shutdown;
  1849. }
  1850. if (tea_shutdown == action)
  1851. queryJobManager().stop();
  1852. }
  1853. }
  1854. return true;
  1855. }
  1856. // CJobMasterChannel
  1857. CJobMasterChannel::CJobMasterChannel(CJobBase &job, IMPServer *mpServer, unsigned channel) : CJobChannel(job, mpServer, channel)
  1858. {
  1859. codeCtx.setown(new CThorCodeContextMaster(*this, job.queryWorkUnit(), job.queryDllEntry(), *job.queryUserDescriptor()));
  1860. sharedMemCodeCtx.setown(new CThorCodeContextMasterSharedMem(*this, job.querySharedAllocator(), job.queryWorkUnit(), job.queryDllEntry(), *job.queryUserDescriptor()));
  1861. }
  1862. CGraphBase *CJobMasterChannel::createGraph()
  1863. {
  1864. return new CMasterGraph(*this);
  1865. }
  1866. IBarrier *CJobMasterChannel::createBarrier(mptag_t tag)
  1867. {
  1868. return new CBarrierMaster(*jobComm, tag);
  1869. }
  1870. ///////////////////
  1871. class CCollatedResult : implements IThorResult, public CSimpleInterface
  1872. {
  1873. CMasterGraph &graph;
  1874. CActivityBase &activity;
  1875. IThorRowInterfaces *rowIf;
  1876. unsigned id;
  1877. CriticalSection crit;
  1878. PointerArrayOf<CThorExpandingRowArray> results;
  1879. Owned<IThorResult> result;
  1880. unsigned spillPriority;
  1881. activity_id ownerId;
  1882. void ensure()
  1883. {
  1884. CriticalBlock b(crit);
  1885. if (result)
  1886. return;
  1887. mptag_t replyTag = graph.queryMPServer().createReplyTag();
  1888. CMessageBuffer msg;
  1889. msg.append(GraphGetResult);
  1890. msg.append(activity.queryJob().queryKey());
  1891. msg.append(graph.queryGraphId());
  1892. msg.append(ownerId);
  1893. msg.append(id);
  1894. msg.append(replyTag);
  1895. ((CJobMaster &)graph.queryJob()).broadcast(queryNodeComm(), msg, masterSlaveMpTag, LONGTIMEOUT, "CCollectResult", NULL, true);
  1896. unsigned numSlaves = graph.queryJob().querySlaves();
  1897. for (unsigned n=0; n<numSlaves; n++)
  1898. results.item(n)->kill();
  1899. rank_t sender;
  1900. MemoryBuffer mb;
  1901. Owned<ISerialStream> stream = createMemoryBufferSerialStream(mb);
  1902. CThorStreamDeserializerSource rowSource(stream);
  1903. unsigned todo = numSlaves;
  1904. for (;;)
  1905. {
  1906. for (;;)
  1907. {
  1908. if (activity.queryAbortSoon())
  1909. return;
  1910. msg.clear();
  1911. if (activity.receiveMsg(msg, RANK_ALL, replyTag, &sender, 60*1000))
  1912. break;
  1913. ActPrintLog(&activity, "WARNING: tag %d timedout, retrying", (unsigned)replyTag);
  1914. }
  1915. sender = sender - 1; // 0 = master
  1916. if (!msg.length())
  1917. {
  1918. --todo;
  1919. if (0 == todo)
  1920. break; // done
  1921. }
  1922. else
  1923. {
  1924. bool error;
  1925. msg.read(error);
  1926. if (error)
  1927. {
  1928. Owned<IThorException> e = deserializeThorException(msg);
  1929. e->setSlave(sender);
  1930. throw e.getClear();
  1931. }
  1932. ThorExpand(msg, mb.clear());
  1933. CThorExpandingRowArray *slaveResults = results.item(sender);
  1934. while (!rowSource.eos())
  1935. {
  1936. RtlDynamicRowBuilder rowBuilder(rowIf->queryRowAllocator());
  1937. size32_t sz = rowIf->queryRowDeserializer()->deserialize(rowBuilder, rowSource);
  1938. slaveResults->append(rowBuilder.finalizeRowClear(sz));
  1939. }
  1940. }
  1941. }
  1942. Owned<IThorResult> _result = ::createResult(activity, rowIf, thorgraphresult_nul, spillPriority);
  1943. Owned<IRowWriter> resultWriter = _result->getWriter();
  1944. for (unsigned s=0; s<numSlaves; s++)
  1945. {
  1946. CThorExpandingRowArray *slaveResult = results.item(s);
  1947. ForEachItemIn(r, *slaveResult)
  1948. {
  1949. const void *row = slaveResult->query(r);
  1950. LinkThorRow(row);
  1951. resultWriter->putRow(row);
  1952. }
  1953. }
  1954. result.setown(_result.getClear());
  1955. }
  1956. public:
  1957. IMPLEMENT_IINTERFACE_USING(CSimpleInterface);
  1958. CCollatedResult(CMasterGraph &_graph, CActivityBase &_activity, IThorRowInterfaces *_rowIf, unsigned _id, activity_id _ownerId, unsigned _spillPriority)
  1959. : graph(_graph), activity(_activity), rowIf(_rowIf), id(_id), ownerId(_ownerId), spillPriority(_spillPriority)
  1960. {
  1961. for (unsigned n=0; n<graph.queryJob().querySlaves(); n++)
  1962. results.append(new CThorExpandingRowArray(activity, rowIf));
  1963. }
  1964. ~CCollatedResult()
  1965. {
  1966. ForEachItemIn(l, results)
  1967. {
  1968. CThorExpandingRowArray *result = results.item(l);
  1969. delete result;
  1970. }
  1971. }
  1972. void setId(unsigned _id)
  1973. {
  1974. id = _id;
  1975. }
  1976. // IThorResult
  1977. virtual IRowWriter *getWriter() { throwUnexpected(); }
  1978. virtual void setResultStream(IRowWriterMultiReader *stream, rowcount_t count)
  1979. {
  1980. throwUnexpected();
  1981. }
  1982. virtual IRowStream *getRowStream()
  1983. {
  1984. ensure();
  1985. return result->getRowStream();
  1986. }
  1987. virtual IThorRowInterfaces *queryRowInterfaces()
  1988. {
  1989. return rowIf;
  1990. }
  1991. virtual CActivityBase *queryActivity()
  1992. {
  1993. return &activity;
  1994. }
  1995. virtual bool isDistributed() const { return false; }
  1996. virtual void serialize(MemoryBuffer &mb)
  1997. {
  1998. ensure();
  1999. result->serialize(mb);
  2000. }
  2001. virtual void getLinkedResult(unsigned & count, const byte * * & ret) override
  2002. {
  2003. ensure();
  2004. result->getLinkedResult(count, ret);
  2005. }
  2006. virtual const void * getLinkedRowResult()
  2007. {
  2008. ensure();
  2009. return result->getLinkedRowResult();
  2010. }
  2011. };
  2012. ///////////////////
  2013. //
  2014. // CMasterGraph impl.
  2015. //
  2016. CMasterGraph::CMasterGraph(CJobChannel &jobChannel) : CGraphBase(jobChannel)
  2017. {
  2018. jobM = (CJobMaster *)&jobChannel.queryJob();
  2019. mpTag = queryJob().allocateMPTag();
  2020. startBarrierTag = queryJob().allocateMPTag();
  2021. waitBarrierTag = queryJob().allocateMPTag();
  2022. startBarrier = jobChannel.createBarrier(startBarrierTag);
  2023. waitBarrier = jobChannel.createBarrier(waitBarrierTag);
  2024. statNumExecutions.setown(new CThorStats(queryJob(), StNumExecutions));
  2025. }
  2026. CMasterGraph::~CMasterGraph()
  2027. {
  2028. queryJob().freeMPTag(mpTag);
  2029. queryJob().freeMPTag(startBarrierTag);
  2030. queryJob().freeMPTag(waitBarrierTag);
  2031. if (TAG_NULL != doneBarrierTag)
  2032. queryJob().freeMPTag(doneBarrierTag);
  2033. if (TAG_NULL != executeReplyTag)
  2034. queryJob().freeMPTag(executeReplyTag);
  2035. }
  2036. void CMasterGraph::init()
  2037. {
  2038. CGraphBase::init();
  2039. if (queryOwner() && isGlobal())
  2040. {
  2041. doneBarrierTag = queryJob().allocateMPTag();
  2042. doneBarrier = jobChannel.createBarrier(doneBarrierTag);
  2043. }
  2044. }
  2045. bool CMasterGraph::fireException(IException *e)
  2046. {
  2047. IThorException *te = QUERYINTERFACE(e, IThorException);
  2048. ThorExceptionAction action;
  2049. if (!te) action = tea_null;
  2050. else action = te->queryAction();
  2051. if (QUERYINTERFACE(e, IMP_Exception) && MPERR_link_closed==e->errorCode())
  2052. action = tea_shutdown;
  2053. else if (QUERYINTERFACE(e, ISEH_Exception))
  2054. action = tea_shutdown;
  2055. CriticalBlock b(exceptCrit);
  2056. switch (action)
  2057. {
  2058. case tea_warning:
  2059. {
  2060. LOG(MCwarning, thorJob, e);
  2061. ErrorSeverity mappedSeverity = job.queryWorkUnit().getWarningSeverity(e->errorCode(), SeverityWarning);
  2062. if (mappedSeverity != SeverityIgnore)
  2063. reportExceptionToWorkunit(job.queryWorkUnit(), e);
  2064. break;
  2065. }
  2066. default:
  2067. {
  2068. LOG(MCerror, thorJob, e);
  2069. if (NULL != fatalHandler)
  2070. fatalHandler->inform(LINK(e));
  2071. if (owner)
  2072. owner->fireException(e);
  2073. else
  2074. queryJobChannel().fireException(e);
  2075. }
  2076. }
  2077. return true;
  2078. }
  2079. void CMasterGraph::reset()
  2080. {
  2081. CGraphBase::reset();
  2082. bcastMsgHandler.reset();
  2083. activityInitMsgHandler.reset();
  2084. executeReplyMsgHandler.reset();
  2085. }
  2086. void CMasterGraph::abort(IException *e)
  2087. {
  2088. if (aborted) return;
  2089. bool _graphDone = graphDone; // aborting master activities can trigger master graphDone, but want to fire GraphAbort to slaves if graphDone=false at start.
  2090. bool dumpInfo = TE_WorkUnitAbortingDumpInfo == e->errorCode() || job.getOptBool("dumpInfoOnAbort");
  2091. if (dumpInfo)
  2092. {
  2093. StringBuffer dumpInfoCmd;
  2094. checkAndDumpAbortInfo(queryJob().getOpt("dumpInfoCmd", dumpInfoCmd));
  2095. }
  2096. try { CGraphBase::abort(e); }
  2097. catch (IException *e)
  2098. {
  2099. GraphPrintLog(e, "Aborting master graph");
  2100. e->Release();
  2101. }
  2102. bcastMsgHandler.cancel(0);
  2103. activityInitMsgHandler.cancel(RANK_ALL);
  2104. executeReplyMsgHandler.cancel(RANK_ALL);
  2105. if (started && !_graphDone)
  2106. {
  2107. try
  2108. {
  2109. CMessageBuffer msg;
  2110. msg.append(GraphAbort);
  2111. msg.append(job.queryKey());
  2112. msg.append(dumpInfo);
  2113. msg.append(queryGraphId());
  2114. jobM->broadcast(queryNodeComm(), msg, masterSlaveMpTag, LONGTIMEOUT, "abort");
  2115. }
  2116. catch (IException *e)
  2117. {
  2118. GraphPrintLog(e, "Aborting slave graph");
  2119. if (abortException)
  2120. {
  2121. e->Release();
  2122. throw LINK(abortException);
  2123. }
  2124. throw;
  2125. }
  2126. }
  2127. if (!queryOwner())
  2128. {
  2129. if (globals->getPropBool("@watchdogProgressEnabled"))
  2130. queryJobManager().queryDeMonServer()->endGraph(this, !queryAborted());
  2131. }
  2132. }
  2133. bool CMasterGraph::serializeActivityInitData(unsigned slave, MemoryBuffer &mb, IThorActivityIterator &iter)
  2134. {
  2135. CriticalBlock b(createdCrit);
  2136. DelayedSizeMarker sizeMark1(mb);
  2137. ForEach (iter)
  2138. {
  2139. CMasterGraphElement &element = (CMasterGraphElement &)iter.query();
  2140. if (!element.sentActInitData->testSet(slave))
  2141. {
  2142. CMasterActivity *activity = (CMasterActivity *)element.queryActivity();
  2143. if (activity)
  2144. {
  2145. mb.append(element.queryId());
  2146. DelayedSizeMarker sizeMark2(mb);
  2147. activity->serializeSlaveData(mb, slave);
  2148. sizeMark2.write();
  2149. }
  2150. }
  2151. }
  2152. if (0 == sizeMark1.size())
  2153. return false;
  2154. mb.append((activity_id)0); // terminator
  2155. sizeMark1.write();
  2156. return true;
  2157. }
  2158. void CMasterGraph::readActivityInitData(MemoryBuffer &mb, unsigned slave)
  2159. {
  2160. for (;;)
  2161. {
  2162. activity_id id;
  2163. mb.read(id);
  2164. if (0 == id)
  2165. break;
  2166. size32_t dataLen;
  2167. mb.read(dataLen);
  2168. CGraphElementBase *element = queryElement(id);
  2169. MemoryBuffer &mbDst = element->queryActivity()->queryInitializationData(slave);
  2170. mbDst.append(dataLen, mb.readDirect(dataLen));
  2171. }
  2172. }
  2173. bool CMasterGraph::prepare(size32_t parentExtractSz, const byte *parentExtract, bool checkDependencies, bool shortCircuit, bool async)
  2174. {
  2175. if (!CGraphBase::prepare(parentExtractSz, parentExtract, checkDependencies, shortCircuit, async)) return false;
  2176. if (aborted) return false;
  2177. return true;
  2178. }
  2179. void CMasterGraph::execute(size32_t _parentExtractSz, const byte *parentExtract, bool checkDependencies, bool async)
  2180. {
  2181. if (isComplete())
  2182. return;
  2183. if (!queryOwner()) // owning graph sends query+child graphs
  2184. jobM->sendQuery(); // if not previously sent
  2185. CGraphBase::execute(parentExtractSz, parentExtract, checkDependencies, async);
  2186. }
  2187. void CMasterGraph::start()
  2188. {
  2189. if (!queryOwner())
  2190. {
  2191. if (globals->getPropBool("@watchdogProgressEnabled"))
  2192. queryJobManager().queryDeMonServer()->startGraph(this);
  2193. }
  2194. Owned<IThorActivityIterator> iter = getConnectedIterator();
  2195. ForEach (*iter)
  2196. iter->query().queryActivity()->startProcess();
  2197. }
  2198. void CMasterGraph::sendActivityInitData()
  2199. {
  2200. CMessageBuffer msg;
  2201. mptag_t replyTag = queryJobChannel().queryMPServer().createReplyTag();
  2202. msg.setReplyTag(replyTag);
  2203. unsigned pos = msg.length();
  2204. unsigned w=0;
  2205. unsigned sentTo = 0;
  2206. Owned<IThorActivityIterator> iter = getConnectedIterator();
  2207. for (; w<queryJob().querySlaves(); w++)
  2208. {
  2209. unsigned needActInit = 0;
  2210. ForEach(*iter)
  2211. {
  2212. CGraphElementBase &element = iter->query();
  2213. CActivityBase *activity = element.queryActivity();
  2214. if (activity && activity->needReInit())
  2215. element.sentActInitData->set(w, false); // force act init to be resent
  2216. if (!element.sentActInitData->test(w)) // has it been sent
  2217. ++needActInit;
  2218. }
  2219. if (needActInit)
  2220. {
  2221. try
  2222. {
  2223. msg.rewrite(pos);
  2224. serializeActivityInitData(w, msg, *iter);
  2225. }
  2226. catch (IException *e)
  2227. {
  2228. GraphPrintLog(e);
  2229. throw;
  2230. }
  2231. if (!queryJobChannel().queryJobComm().send(msg, w+1, mpTag, LONGTIMEOUT))
  2232. {
  2233. StringBuffer epStr;
  2234. throw MakeStringException(0, "Timeout sending to slave %s", job.querySlaveGroup().queryNode(w).endpoint().getUrlStr(epStr).str());
  2235. }
  2236. ++sentTo;
  2237. }
  2238. }
  2239. if (sentTo)
  2240. {
  2241. assertex(sentTo == queryJob().querySlaves());
  2242. w=0;
  2243. Owned<IException> e;
  2244. // now get back initialization data from graph tag
  2245. for (; w<queryJob().querySlaves(); w++)
  2246. {
  2247. rank_t sender;
  2248. msg.clear();
  2249. if (!activityInitMsgHandler.recv(queryJobChannel().queryJobComm(), msg, w+1, replyTag, &sender, LONGTIMEOUT))
  2250. throw MakeGraphException(this, 0, "Timeout receiving from slaves after graph sent");
  2251. bool error;
  2252. msg.read(error);
  2253. if (error)
  2254. {
  2255. Owned<IThorException> se = deserializeThorException(msg);
  2256. se->setSlave(sender);
  2257. if (!e.get())
  2258. {
  2259. StringBuffer tmpStr("Slave ");
  2260. queryJob().queryJobGroup().queryNode(sender).endpoint().getUrlStr(tmpStr);
  2261. GraphPrintLog(se, "%s", tmpStr.append(": slave initialization error").str());
  2262. e.setown(se.getClear());
  2263. }
  2264. continue; // to read other slave responses.
  2265. }
  2266. readActivityInitData(msg, w);
  2267. }
  2268. if (e.get())
  2269. throw LINK(e);
  2270. }
  2271. }
  2272. void CMasterGraph::serializeGraphInit(MemoryBuffer &mb)
  2273. {
  2274. mb.append(graphId);
  2275. serializeMPtag(mb, mpTag);
  2276. mb.append((int)startBarrierTag);
  2277. mb.append((int)waitBarrierTag);
  2278. mb.append((int)doneBarrierTag);
  2279. mb.append(queryChildGraphCount());
  2280. Owned<IThorGraphIterator> childIter = getChildGraphIterator();
  2281. ForEach (*childIter)
  2282. {
  2283. CMasterGraph &childGraph = (CMasterGraph &)childIter->query();
  2284. childGraph.serializeGraphInit(mb);
  2285. }
  2286. }
  2287. // IThorChildGraph impl.
  2288. IEclGraphResults *CMasterGraph::evaluate(unsigned _parentExtractSz, const byte *parentExtract)
  2289. {
  2290. throw MakeGraphException(this, 0, "Thor master does not support the execution of child queries");
  2291. }
  2292. void CMasterGraph::executeSubGraph(size32_t parentExtractSz, const byte *parentExtract)
  2293. {
  2294. if (job.queryResumed()) // skip complete subgraph if resuming. NB: temp spill have been tucked away for this purpose when paused.
  2295. {
  2296. if (!queryOwner())
  2297. {
  2298. if (WUGraphComplete == job.queryWorkUnit().queryNodeState(job.queryGraphName(), graphId))
  2299. setCompleteEx();
  2300. }
  2301. }
  2302. if (isComplete())
  2303. return;
  2304. if (!sentGlobalInit)
  2305. {
  2306. sentGlobalInit = true;
  2307. if (!queryOwner())
  2308. sendGraph();
  2309. else
  2310. {
  2311. if (isGlobal())
  2312. {
  2313. CMessageBuffer msg;
  2314. serializeCreateContexts(msg);
  2315. try
  2316. {
  2317. jobM->broadcast(queryJobChannel().queryJobComm(), msg, mpTag, LONGTIMEOUT, "serializeCreateContexts", &bcastMsgHandler);
  2318. }
  2319. catch (IException *e)
  2320. {
  2321. GraphPrintLog(e, "Aborting graph create(2)");
  2322. if (abortException)
  2323. {
  2324. e->Release();
  2325. throw LINK(abortException);
  2326. }
  2327. throw;
  2328. }
  2329. }
  2330. }
  2331. }
  2332. fatalHandler.clear();
  2333. fatalHandler.setown(new CFatalHandler(globals->getPropInt("@fatal_timeout", FATAL_TIMEOUT)));
  2334. CGraphBase::executeSubGraph(parentExtractSz, parentExtract);
  2335. if (TAG_NULL != executeReplyTag)
  2336. {
  2337. rank_t sender;
  2338. unsigned s=0;
  2339. for (; s<queryJob().querySlaves(); s++)
  2340. {
  2341. CMessageBuffer msg;
  2342. if (!executeReplyMsgHandler.recv(queryJobChannel().queryJobComm(), msg, RANK_ALL, executeReplyTag, &sender))
  2343. break;
  2344. bool error;
  2345. msg.read(error);
  2346. if (error)
  2347. {
  2348. Owned<IThorException> exception = deserializeThorException(msg);
  2349. exception->setSlave(sender);
  2350. GraphPrintLog(exception, "slave execute reply exception");
  2351. throw exception.getClear();
  2352. }
  2353. }
  2354. if (fatalHandler)
  2355. fatalHandler->clear();
  2356. }
  2357. fatalHandler.clear();
  2358. Owned<IWorkUnit> wu = &job.queryWorkUnit().lock();
  2359. queryJobManager().updateWorkUnitLog(*wu);
  2360. }
  2361. void CMasterGraph::sendGraph()
  2362. {
  2363. CTimeMon atimer;
  2364. CMessageBuffer msg;
  2365. msg.append(GraphInit);
  2366. msg.append(job.queryKey());
  2367. if (TAG_NULL == executeReplyTag)
  2368. executeReplyTag = jobM->allocateMPTag();
  2369. serializeMPtag(msg, executeReplyTag);
  2370. serializeCreateContexts(msg);
  2371. serializeGraphInit(msg);
  2372. // slave graph data
  2373. try
  2374. {
  2375. jobM->broadcast(queryNodeComm(), msg, masterSlaveMpTag, LONGTIMEOUT, "sendGraph", &bcastMsgHandler);
  2376. }
  2377. catch (IException *e)
  2378. {
  2379. GraphPrintLog(e, "Aborting sendGraph");
  2380. if (abortException)
  2381. {
  2382. e->Release();
  2383. throw LINK(abortException);
  2384. }
  2385. throw;
  2386. }
  2387. GraphPrintLog("sendGraph took %d ms", atimer.elapsed());
  2388. }
  2389. bool CMasterGraph::preStart(size32_t parentExtractSz, const byte *parentExtract)
  2390. {
  2391. GraphPrintLog("Processing graph");
  2392. if (syncInitData())
  2393. {
  2394. sendActivityInitData(); // has to be done at least once
  2395. // NB: At this point, on the slaves, the graphs will start
  2396. }
  2397. CGraphBase::preStart(parentExtractSz, parentExtract);
  2398. if (isGlobal())
  2399. {
  2400. if (!startBarrier->wait(true)) // ensure all graphs are at start point at same time, as some may request data from each other.
  2401. return false;
  2402. }
  2403. if (!queryOwner())
  2404. job.queryWorkUnit().setNodeState(job.queryGraphName(), graphId, WUGraphRunning);
  2405. return true;
  2406. }
  2407. void CMasterGraph::handleSlaveDone(unsigned node, MemoryBuffer &mb)
  2408. {
  2409. graph_id gid;
  2410. mb.read(gid);
  2411. assertex(gid == queryGraphId());
  2412. unsigned count;
  2413. mb.read(count);
  2414. while (count--)
  2415. {
  2416. activity_id activityId;
  2417. mb.read(activityId);
  2418. CMasterGraphElement *act = (CMasterGraphElement *)queryElement(activityId);
  2419. unsigned len;
  2420. mb.read(len);
  2421. const void *d = mb.readDirect(len);
  2422. MemoryBuffer sdMb;
  2423. sdMb.setBuffer(len, (void *)d);
  2424. act->slaveDone(node, sdMb);
  2425. }
  2426. }
  2427. void CMasterGraph::getFinalProgress()
  2428. {
  2429. offset_t totalDiskUsage = 0;
  2430. offset_t minNodeDiskUsage = 0, maxNodeDiskUsage = 0;
  2431. unsigned maxNode = (unsigned)-1, minNode = (unsigned)-1;
  2432. CMessageBuffer msg;
  2433. mptag_t replyTag = queryJobChannel().queryMPServer().createReplyTag();
  2434. msg.setReplyTag(replyTag);
  2435. msg.append((unsigned)GraphEnd);
  2436. msg.append(job.queryKey());
  2437. msg.append(queryGraphId());
  2438. jobM->broadcast(queryNodeComm(), msg, masterSlaveMpTag, LONGTIMEOUT, "graphEnd", NULL, true);
  2439. Owned<IBitSet> respondedBitSet = createBitSet();
  2440. unsigned n=queryJob().queryNodes();
  2441. while (n--)
  2442. {
  2443. rank_t sender;
  2444. if (!queryNodeComm().recv(msg, RANK_ALL, replyTag, &sender, LONGTIMEOUT))
  2445. {
  2446. StringBuffer slaveList;
  2447. n=queryJob().queryNodes();
  2448. unsigned s = 0;
  2449. for (;;)
  2450. {
  2451. unsigned ns = respondedBitSet->scan(s, true); // scan for next respondent
  2452. // list all non-respondents up to this found respondent (or rest of slaves)
  2453. while (s<ns && s<n) // NB: if scan find nothing else set, ns = (unsigned)-1
  2454. {
  2455. if (slaveList.length())
  2456. slaveList.append(",");
  2457. ++s; // inc. 1st as slaveList is list of slaves that are 1 based.
  2458. slaveList.append(s);
  2459. }
  2460. ++s;
  2461. if (s>=n)
  2462. break;
  2463. }
  2464. throw MakeGraphException(this, 0, "Timeout receiving final progress from slaves - these slaves failed to respond: %s", slaveList.str());
  2465. }
  2466. bool error;
  2467. msg.read(error);
  2468. if (error)
  2469. {
  2470. Owned<IThorException> e = deserializeThorException(msg);
  2471. e->setSlave(sender);
  2472. throw e.getClear();
  2473. }
  2474. respondedBitSet->set(((unsigned)sender)-1);
  2475. if (0 == msg.remaining())
  2476. continue;
  2477. unsigned channelsPerSlave = globals->getPropInt("@channelsPerSlave", 1); // JCSMORE - should move somewhere common
  2478. for (unsigned sc=0; sc<channelsPerSlave; sc++)
  2479. {
  2480. unsigned slave;
  2481. msg.read(slave);
  2482. handleSlaveDone(slave, msg);
  2483. if (!queryOwner())
  2484. {
  2485. if (globals->getPropBool("@watchdogProgressEnabled"))
  2486. {
  2487. try
  2488. {
  2489. size32_t progressLen;
  2490. msg.read(progressLen);
  2491. MemoryBuffer progressData;
  2492. progressData.setBuffer(progressLen, (void *)msg.readDirect(progressLen));
  2493. queryJobManager().queryDeMonServer()->takeHeartBeat(progressData);
  2494. }
  2495. catch (IException *e)
  2496. {
  2497. GraphPrintLog(e, "Failure whilst deserializing stats/progress");
  2498. e->Release();
  2499. }
  2500. }
  2501. }
  2502. offset_t nodeDiskUsage;
  2503. msg.read(nodeDiskUsage);
  2504. jobM->setNodeDiskUsage(n, nodeDiskUsage);
  2505. if (nodeDiskUsage > maxNodeDiskUsage)
  2506. {
  2507. maxNodeDiskUsage = nodeDiskUsage;
  2508. maxNode = n;
  2509. }
  2510. if ((unsigned)-1 == minNode || nodeDiskUsage < minNodeDiskUsage)
  2511. {
  2512. minNodeDiskUsage = nodeDiskUsage;
  2513. minNode = n;
  2514. }
  2515. totalDiskUsage += nodeDiskUsage;
  2516. }
  2517. }
  2518. if (totalDiskUsage)
  2519. {
  2520. Owned<IWorkUnit> wu = &job.queryWorkUnit().lock();
  2521. wu->addDiskUsageStats(totalDiskUsage/queryJob().querySlaves(), minNode, minNodeDiskUsage, maxNode, maxNodeDiskUsage, queryGraphId());
  2522. }
  2523. }
  2524. void CMasterGraph::done()
  2525. {
  2526. if (started)
  2527. {
  2528. if (!aborted && (!queryOwner() || isGlobal()))
  2529. getFinalProgress(); // waiting for slave graph to finish and send final progress update + extra act end info.
  2530. }
  2531. CGraphBase::done();
  2532. if (started && !queryOwner())
  2533. {
  2534. if (globals->getPropBool("@watchdogProgressEnabled"))
  2535. queryJobManager().queryDeMonServer()->endGraph(this, true);
  2536. }
  2537. }
  2538. void CMasterGraph::setComplete(bool tf)
  2539. {
  2540. CGraphBase::setComplete(tf);
  2541. if (tf && !queryOwner())
  2542. {
  2543. job.queryWorkUnit().setNodeState(job.queryGraphName(), graphId, graphDone?WUGraphComplete:WUGraphFailed);
  2544. }
  2545. }
  2546. bool CMasterGraph::deserializeStats(unsigned node, MemoryBuffer &mb)
  2547. {
  2548. CriticalBlock b(createdCrit);
  2549. unsigned numExecutions;
  2550. mb.read(numExecutions);
  2551. statNumExecutions->set(node, numExecutions);
  2552. unsigned count;
  2553. mb.read(count);
  2554. if (count)
  2555. setProgressUpdated();
  2556. while (count--)
  2557. {
  2558. activity_id activityId;
  2559. mb.read(activityId);
  2560. CMasterActivity *activity = NULL;
  2561. CMasterGraphElement *element = (CMasterGraphElement *)queryElement(activityId);
  2562. if (element)
  2563. {
  2564. activity = (CMasterActivity *)element->queryActivity();
  2565. if (!activity)
  2566. {
  2567. CGraphBase *parentGraph = element->queryOwner().queryOwner(); // i.e. am I in a childgraph
  2568. if (!parentGraph)
  2569. {
  2570. GraphPrintLog("Activity id=%" ACTPF "d not created in master and not a child query activity", activityId);
  2571. return false; // don't know if or how this could happen, but all bets off with packet if did.
  2572. }
  2573. Owned<IException> e;
  2574. try
  2575. {
  2576. element->createActivity();
  2577. activity = (CMasterActivity *)element->queryActivity();
  2578. }
  2579. catch (IException *_e)
  2580. {
  2581. e.setown(_e);
  2582. GraphPrintLog(_e, "In deserializeStats");
  2583. }
  2584. if (!activity || e.get())
  2585. {
  2586. GraphPrintLog("Activity id=%" ACTPF "d failed to created child query activity ready for progress", activityId);
  2587. return false;
  2588. }
  2589. }
  2590. if (activity)
  2591. activity->deserializeStats(node, mb);
  2592. }
  2593. else
  2594. {
  2595. GraphPrintLog("Failed to find activity, during progress deserialization, id=%" ACTPF "d", activityId);
  2596. return false; // don't know if or how this could happen, but all bets off with packet if did.
  2597. }
  2598. }
  2599. unsigned subs, _subs;
  2600. mb.read(subs);
  2601. _subs = subs;
  2602. while (subs--)
  2603. {
  2604. graph_id subId;
  2605. mb.read(subId);
  2606. Owned<CMasterGraph> graph = (CMasterGraph *)job.queryJobChannel(0).getGraph(subId);
  2607. if (NULL == graph.get())
  2608. return false;
  2609. if (!graph->deserializeStats(node, mb))
  2610. return false;
  2611. }
  2612. return true;
  2613. }
  2614. void CMasterGraph::getStats(IStatisticGatherer &stats)
  2615. {
  2616. // graph specific stats
  2617. statNumExecutions->getStats(stats, false);
  2618. Owned<IThorActivityIterator> iter;
  2619. if (queryOwner() && !isGlobal())
  2620. iter.setown(getIterator()); // Local child graphs still send progress, but aren't connected in master
  2621. else
  2622. iter.setown(getConnectedIterator());
  2623. ForEach (*iter)
  2624. {
  2625. CMasterGraphElement &container = (CMasterGraphElement &)iter->query();
  2626. CMasterActivity *activity = (CMasterActivity *)container.queryActivity();
  2627. if (activity) // may not be created (if within child query)
  2628. {
  2629. activity_id id = container.queryId();
  2630. unsigned outputs = container.getOutputs();
  2631. unsigned oid = 0;
  2632. for (; oid < outputs; oid++)
  2633. {
  2634. StatsEdgeScope edgeScope(stats, id, oid);
  2635. activity->getEdgeStats(stats, oid); // for subgraph, may recursively call reportGraph
  2636. }
  2637. StatsActivityScope scope(stats, id);
  2638. activity->getActivityStats(stats);
  2639. }
  2640. }
  2641. }
  2642. IThorResult *CMasterGraph::createResult(CActivityBase &activity, unsigned id, IThorGraphResults *results, IThorRowInterfaces *rowIf, ThorGraphResultType resultType, unsigned spillPriority)
  2643. {
  2644. Owned<CCollatedResult> result = new CCollatedResult(*this, activity, rowIf, id, results->queryOwnerId(), spillPriority);
  2645. results->setResult(id, result);
  2646. return result;
  2647. }
  2648. IThorResult *CMasterGraph::createResult(CActivityBase &activity, unsigned id, IThorRowInterfaces *rowIf, ThorGraphResultType resultType, unsigned spillPriority)
  2649. {
  2650. Owned<CCollatedResult> result = new CCollatedResult(*this, activity, rowIf, id, localResults->queryOwnerId(), spillPriority);
  2651. localResults->setResult(id, result);
  2652. return result;
  2653. }
  2654. IThorResult *CMasterGraph::createGraphLoopResult(CActivityBase &activity, IThorRowInterfaces *rowIf, ThorGraphResultType resultType, unsigned spillPriority)
  2655. {
  2656. Owned<CCollatedResult> result = new CCollatedResult(*this, activity, rowIf, 0, localResults->queryOwnerId(), spillPriority);
  2657. unsigned id = graphLoopResults->addResult(result);
  2658. result->setId(id);
  2659. return result;
  2660. }
  2661. ///////////////////////////////////////////////////
  2662. static bool suppressStatisticIfZero(StatisticKind kind)
  2663. {
  2664. switch (kind)
  2665. {
  2666. case StNumSpills:
  2667. case StSizeSpillFile:
  2668. case StTimeSpillElapsed:
  2669. case StNumDiskRetries:
  2670. return true;
  2671. }
  2672. return false;
  2673. }
  2674. ///////////////////////////////////////////////////
  2675. CThorStats::CThorStats(CJobBase &_ctx, StatisticKind _kind) : ctx(_ctx), kind(_kind)
  2676. {
  2677. unsigned c = queryClusterWidth();
  2678. while (c--) counts.append(0);
  2679. reset();
  2680. }
  2681. void CThorStats::extract(unsigned node, const CRuntimeStatisticCollection & stats)
  2682. {
  2683. set(node, stats.getStatisticValue(kind));
  2684. }
  2685. void CThorStats::set(unsigned node, unsigned __int64 count)
  2686. {
  2687. counts.replace(count, node);
  2688. }
  2689. void CThorStats::reset()
  2690. {
  2691. tot = max = avg = 0;
  2692. min = (unsigned __int64) -1;
  2693. minNode = maxNode = maxSkew = minSkew = maxNode = minNode = 0;
  2694. }
  2695. void CThorStats::calculateSkew()
  2696. {
  2697. unsigned count = counts.ordinality();
  2698. double _avg = (double)tot/count;
  2699. if (_avg)
  2700. {
  2701. if (max > ctx.querySlaves()) // i.e. if small count, suppress skew stats.
  2702. {
  2703. //MORE: Range protection on maxSkew?
  2704. maxSkew = (unsigned)(10000.0 * (((double)max-_avg)/_avg));
  2705. minSkew = (unsigned)(10000.0 * ((_avg-(double)min)/_avg));
  2706. }
  2707. avg = (unsigned __int64)_avg;
  2708. }
  2709. }
  2710. void CThorStats::tallyValue(unsigned __int64 thiscount, unsigned n)
  2711. {
  2712. tot += thiscount;
  2713. if (thiscount > max)
  2714. {
  2715. max = thiscount;
  2716. maxNode = n;
  2717. }
  2718. if (thiscount < min)
  2719. {
  2720. min = thiscount;
  2721. minNode = n;
  2722. }
  2723. }
  2724. void CThorStats::processTotal()
  2725. {
  2726. reset();
  2727. ForEachItemIn(n, counts)
  2728. {
  2729. unsigned __int64 thiscount = counts.item(n);
  2730. tallyValue(thiscount, n+1);
  2731. }
  2732. }
  2733. void CThorStats::processInfo()
  2734. {
  2735. processTotal();
  2736. calculateSkew();
  2737. }
  2738. void CThorStats::getTotalStat(IStatisticGatherer & stats)
  2739. {
  2740. processTotal();
  2741. stats.addStatistic(kind, tot);
  2742. }
  2743. void CThorStats::getStats(IStatisticGatherer & stats, bool suppressMinMaxWhenEqual)
  2744. {
  2745. processInfo();
  2746. if ((0 == tot) && suppressStatisticIfZero(kind))
  2747. return;
  2748. //MORE: For most measures (not time stamps etc.) it would be sensible to output the total here....
  2749. if (!suppressMinMaxWhenEqual || (maxSkew != minSkew))
  2750. {
  2751. stats.addStatistic((StatisticKind)(kind|StMinX), min);
  2752. stats.addStatistic((StatisticKind)(kind|StMaxX), max);
  2753. stats.addStatistic((StatisticKind)(kind|StAvgX), avg);
  2754. }
  2755. if (maxSkew != minSkew)
  2756. {
  2757. stats.addStatistic((StatisticKind)(kind|StSkewMin), -(__int64)minSkew); // Save minimum as a negative value so consistent
  2758. stats.addStatistic((StatisticKind)(kind|StSkewMax), maxSkew);
  2759. stats.addStatistic((StatisticKind)(kind|StNodeMin), minNode);
  2760. stats.addStatistic((StatisticKind)(kind|StNodeMax), maxNode);
  2761. }
  2762. }
  2763. ///////////////////////////////////////////////////
  2764. CTimingInfo::CTimingInfo(CJobBase &ctx) : CThorStats(ctx, StTimeLocalExecute)
  2765. {
  2766. }
  2767. ///////////////////////////////////////////////////
  2768. ProgressInfo::ProgressInfo(CJobBase &ctx) : CThorStats(ctx, StNumRowsProcessed)
  2769. {
  2770. startcount = stopcount = 0;
  2771. }
  2772. void ProgressInfo::processInfo() // reimplement as counts have special flags (i.e. stop/start)
  2773. {
  2774. reset();
  2775. startcount = stopcount = 0;
  2776. ForEachItemIn(n, counts)
  2777. {
  2778. unsigned __int64 thiscount = counts.item(n);
  2779. if (thiscount & THORDATALINK_STARTED)
  2780. startcount++;
  2781. if (thiscount & THORDATALINK_STOPPED)
  2782. stopcount++;
  2783. thiscount = thiscount & THORDATALINK_COUNT_MASK;
  2784. tallyValue(thiscount, n+1);
  2785. }
  2786. calculateSkew();
  2787. }
  2788. void ProgressInfo::getStats(IStatisticGatherer & stats)
  2789. {
  2790. CThorStats::getStats(stats, true);
  2791. stats.addStatistic(kind, tot);
  2792. stats.addStatistic(StNumSlaves, counts.ordinality());
  2793. stats.addStatistic(StNumStarts, startcount);
  2794. stats.addStatistic(StNumStops, stopcount);
  2795. }
  2796. ///////////////////////////////////////////////////
  2797. CJobMaster *createThorGraph(const char *graphName, IConstWorkUnit &workunit, ILoadedDllEntry *querySo, bool sendSo, const SocketEndpoint &agentEp)
  2798. {
  2799. Owned<CJobMaster> jobMaster = new CJobMaster(workunit, graphName, querySo, sendSo, agentEp);
  2800. IPropertyTree *graphXGMML = jobMaster->queryGraphXGMML();
  2801. Owned<IPropertyTreeIterator> iter = graphXGMML->getElements("node");
  2802. ForEach(*iter)
  2803. jobMaster->addSubGraph(iter->query());
  2804. jobMaster->addDependencies(graphXGMML);
  2805. return LINK(jobMaster);
  2806. }
  2807. static IJobManager *jobManager = NULL;
  2808. void setJobManager(IJobManager *_jobManager)
  2809. {
  2810. CriticalBlock b(*jobManagerCrit);
  2811. jobManager = _jobManager;
  2812. }
  2813. IJobManager *getJobManager()
  2814. {
  2815. CriticalBlock b(*jobManagerCrit);
  2816. return LINK(jobManager);
  2817. }
  2818. IJobManager &queryJobManager()
  2819. {
  2820. CriticalBlock b(*jobManagerCrit);
  2821. assertex(jobManager);
  2822. return *jobManager;
  2823. }