thgraphmaster.cpp 100 KB

1234567891011121314151617181920212223242526272829303132333435363738394041424344454647484950515253545556575859606162636465666768697071727374757677787980818283848586878889909192939495969798991001011021031041051061071081091101111121131141151161171181191201211221231241251261271281291301311321331341351361371381391401411421431441451461471481491501511521531541551561571581591601611621631641651661671681691701711721731741751761771781791801811821831841851861871881891901911921931941951961971981992002012022032042052062072082092102112122132142152162172182192202212222232242252262272282292302312322332342352362372382392402412422432442452462472482492502512522532542552562572582592602612622632642652662672682692702712722732742752762772782792802812822832842852862872882892902912922932942952962972982993003013023033043053063073083093103113123133143153163173183193203213223233243253263273283293303313323333343353363373383393403413423433443453463473483493503513523533543553563573583593603613623633643653663673683693703713723733743753763773783793803813823833843853863873883893903913923933943953963973983994004014024034044054064074084094104114124134144154164174184194204214224234244254264274284294304314324334344354364374384394404414424434444454464474484494504514524534544554564574584594604614624634644654664674684694704714724734744754764774784794804814824834844854864874884894904914924934944954964974984995005015025035045055065075085095105115125135145155165175185195205215225235245255265275285295305315325335345355365375385395405415425435445455465475485495505515525535545555565575585595605615625635645655665675685695705715725735745755765775785795805815825835845855865875885895905915925935945955965975985996006016026036046056066076086096106116126136146156166176186196206216226236246256266276286296306316326336346356366376386396406416426436446456466476486496506516526536546556566576586596606616626636646656666676686696706716726736746756766776786796806816826836846856866876886896906916926936946956966976986997007017027037047057067077087097107117127137147157167177187197207217227237247257267277287297307317327337347357367377387397407417427437447457467477487497507517527537547557567577587597607617627637647657667677687697707717727737747757767777787797807817827837847857867877887897907917927937947957967977987998008018028038048058068078088098108118128138148158168178188198208218228238248258268278288298308318328338348358368378388398408418428438448458468478488498508518528538548558568578588598608618628638648658668678688698708718728738748758768778788798808818828838848858868878888898908918928938948958968978988999009019029039049059069079089099109119129139149159169179189199209219229239249259269279289299309319329339349359369379389399409419429439449459469479489499509519529539549559569579589599609619629639649659669679689699709719729739749759769779789799809819829839849859869879889899909919929939949959969979989991000100110021003100410051006100710081009101010111012101310141015101610171018101910201021102210231024102510261027102810291030103110321033103410351036103710381039104010411042104310441045104610471048104910501051105210531054105510561057105810591060106110621063106410651066106710681069107010711072107310741075107610771078107910801081108210831084108510861087108810891090109110921093109410951096109710981099110011011102110311041105110611071108110911101111111211131114111511161117111811191120112111221123112411251126112711281129113011311132113311341135113611371138113911401141114211431144114511461147114811491150115111521153115411551156115711581159116011611162116311641165116611671168116911701171117211731174117511761177117811791180118111821183118411851186118711881189119011911192119311941195119611971198119912001201120212031204120512061207120812091210121112121213121412151216121712181219122012211222122312241225122612271228122912301231123212331234123512361237123812391240124112421243124412451246124712481249125012511252125312541255125612571258125912601261126212631264126512661267126812691270127112721273127412751276127712781279128012811282128312841285128612871288128912901291129212931294129512961297129812991300130113021303130413051306130713081309131013111312131313141315131613171318131913201321132213231324132513261327132813291330133113321333133413351336133713381339134013411342134313441345134613471348134913501351135213531354135513561357135813591360136113621363136413651366136713681369137013711372137313741375137613771378137913801381138213831384138513861387138813891390139113921393139413951396139713981399140014011402140314041405140614071408140914101411141214131414141514161417141814191420142114221423142414251426142714281429143014311432143314341435143614371438143914401441144214431444144514461447144814491450145114521453145414551456145714581459146014611462146314641465146614671468146914701471147214731474147514761477147814791480148114821483148414851486148714881489149014911492149314941495149614971498149915001501150215031504150515061507150815091510151115121513151415151516151715181519152015211522152315241525152615271528152915301531153215331534153515361537153815391540154115421543154415451546154715481549155015511552155315541555155615571558155915601561156215631564156515661567156815691570157115721573157415751576157715781579158015811582158315841585158615871588158915901591159215931594159515961597159815991600160116021603160416051606160716081609161016111612161316141615161616171618161916201621162216231624162516261627162816291630163116321633163416351636163716381639164016411642164316441645164616471648164916501651165216531654165516561657165816591660166116621663166416651666166716681669167016711672167316741675167616771678167916801681168216831684168516861687168816891690169116921693169416951696169716981699170017011702170317041705170617071708170917101711171217131714171517161717171817191720172117221723172417251726172717281729173017311732173317341735173617371738173917401741174217431744174517461747174817491750175117521753175417551756175717581759176017611762176317641765176617671768176917701771177217731774177517761777177817791780178117821783178417851786178717881789179017911792179317941795179617971798179918001801180218031804180518061807180818091810181118121813181418151816181718181819182018211822182318241825182618271828182918301831183218331834183518361837183818391840184118421843184418451846184718481849185018511852185318541855185618571858185918601861186218631864186518661867186818691870187118721873187418751876187718781879188018811882188318841885188618871888188918901891189218931894189518961897189818991900190119021903190419051906190719081909191019111912191319141915191619171918191919201921192219231924192519261927192819291930193119321933193419351936193719381939194019411942194319441945194619471948194919501951195219531954195519561957195819591960196119621963196419651966196719681969197019711972197319741975197619771978197919801981198219831984198519861987198819891990199119921993199419951996199719981999200020012002200320042005200620072008200920102011201220132014201520162017201820192020202120222023202420252026202720282029203020312032203320342035203620372038203920402041204220432044204520462047204820492050205120522053205420552056205720582059206020612062206320642065206620672068206920702071207220732074207520762077207820792080208120822083208420852086208720882089209020912092209320942095209620972098209921002101210221032104210521062107210821092110211121122113211421152116211721182119212021212122212321242125212621272128212921302131213221332134213521362137213821392140214121422143214421452146214721482149215021512152215321542155215621572158215921602161216221632164216521662167216821692170217121722173217421752176217721782179218021812182218321842185218621872188218921902191219221932194219521962197219821992200220122022203220422052206220722082209221022112212221322142215221622172218221922202221222222232224222522262227222822292230223122322233223422352236223722382239224022412242224322442245224622472248224922502251225222532254225522562257225822592260226122622263226422652266226722682269227022712272227322742275227622772278227922802281228222832284228522862287228822892290229122922293229422952296229722982299230023012302230323042305230623072308230923102311231223132314231523162317231823192320232123222323232423252326232723282329233023312332233323342335233623372338233923402341234223432344234523462347234823492350235123522353235423552356235723582359236023612362236323642365236623672368236923702371237223732374237523762377237823792380238123822383238423852386238723882389239023912392239323942395239623972398239924002401240224032404240524062407240824092410241124122413241424152416241724182419242024212422242324242425242624272428242924302431243224332434243524362437243824392440244124422443244424452446244724482449245024512452245324542455245624572458245924602461246224632464246524662467246824692470247124722473247424752476247724782479248024812482248324842485248624872488248924902491249224932494249524962497249824992500250125022503250425052506250725082509251025112512251325142515251625172518251925202521252225232524252525262527252825292530253125322533253425352536253725382539254025412542254325442545254625472548254925502551255225532554255525562557255825592560256125622563256425652566256725682569257025712572257325742575257625772578257925802581258225832584258525862587258825892590259125922593259425952596259725982599260026012602260326042605260626072608260926102611261226132614261526162617261826192620262126222623262426252626262726282629263026312632263326342635263626372638263926402641264226432644264526462647264826492650265126522653265426552656265726582659266026612662266326642665266626672668266926702671267226732674267526762677267826792680268126822683268426852686268726882689269026912692269326942695269626972698269927002701270227032704270527062707270827092710271127122713271427152716271727182719272027212722272327242725272627272728272927302731273227332734273527362737273827392740274127422743274427452746274727482749275027512752275327542755275627572758275927602761276227632764276527662767276827692770277127722773277427752776277727782779278027812782278327842785278627872788278927902791279227932794279527962797279827992800280128022803280428052806280728082809281028112812281328142815281628172818281928202821282228232824282528262827282828292830283128322833283428352836283728382839284028412842284328442845284628472848284928502851285228532854285528562857285828592860286128622863286428652866286728682869287028712872287328742875287628772878287928802881288228832884288528862887288828892890289128922893289428952896289728982899290029012902290329042905290629072908290929102911291229132914291529162917291829192920292129222923292429252926292729282929293029312932293329342935
  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 && activity->needReInit())
  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. {
  471. if (!notedWarnings->testSet(e->errorCode()))
  472. CActivityBase::fireException(e);
  473. return true;
  474. }
  475. }
  476. return container.queryOwner().fireException(e);
  477. }
  478. void CMasterActivity::reset()
  479. {
  480. asyncStart = false;
  481. CActivityBase::reset();
  482. }
  483. void CMasterActivity::deserializeStats(unsigned node, MemoryBuffer &mb)
  484. {
  485. CriticalBlock b(progressCrit); // don't think needed
  486. unsigned __int64 localTimeNs;
  487. mb.read(localTimeNs);
  488. timingInfo.set(node, localTimeNs);
  489. rowcount_t count;
  490. ForEachItemIn(p, progressInfo)
  491. {
  492. mb.read(count);
  493. progressInfo.item(p).set(node, count);
  494. }
  495. }
  496. void CMasterActivity::getActivityStats(IStatisticGatherer & stats)
  497. {
  498. timingInfo.getStats(stats);
  499. }
  500. void CMasterActivity::getEdgeStats(IStatisticGatherer & stats, unsigned idx)
  501. {
  502. CriticalBlock b(progressCrit);
  503. if (progressInfo.isItem(idx))
  504. progressInfo.item(idx).getStats(stats);
  505. }
  506. void CMasterActivity::done()
  507. {
  508. CActivityBase::done();
  509. ForEachItemIn(s, readFiles)
  510. {
  511. IDistributedFile &file = readFiles.item(s);
  512. file.setAccessed();
  513. }
  514. }
  515. //////////////////////
  516. // CMasterGraphElement impl.
  517. //
  518. CMasterGraphElement::CMasterGraphElement(CGraphBase &_owner, IPropertyTree &_xgmml) : CGraphElementBase(_owner, _xgmml)
  519. {
  520. }
  521. bool CMasterGraphElement::checkUpdate()
  522. {
  523. if (!onlyUpdateIfChanged)
  524. return false;
  525. if (!globals->getPropBool("@updateEnabled", true) || 0 != queryJob().getWorkUnitValueInt("disableUpdate", 0))
  526. return false;
  527. bool doCheckUpdate = false;
  528. OwnedRoxieString filename;
  529. unsigned eclCRC;
  530. unsigned __int64 totalCRC;
  531. bool temporary = false;
  532. switch (getKind())
  533. {
  534. case TAKindexwrite:
  535. {
  536. IHThorIndexWriteArg *helper = (IHThorIndexWriteArg *)queryHelper();
  537. doCheckUpdate = 0 != (helper->getFlags() & TIWupdate);
  538. filename.set(helper->getFileName());
  539. helper->getUpdateCRCs(eclCRC, totalCRC);
  540. break;
  541. }
  542. case TAKdiskwrite:
  543. case TAKcsvwrite:
  544. case TAKxmlwrite:
  545. case TAKjsonwrite:
  546. case TAKspillwrite:
  547. {
  548. IHThorDiskWriteArg *helper = (IHThorDiskWriteArg *)queryHelper();
  549. doCheckUpdate = 0 != (helper->getFlags() & TDWupdate);
  550. filename.set(helper->getFileName());
  551. helper->getUpdateCRCs(eclCRC, totalCRC);
  552. if (TAKdiskwrite == getKind())
  553. temporary = 0 != (helper->getFlags() & (TDXtemporary|TDXjobtemp));
  554. else if (TAKspillwrite == getKind())
  555. temporary = true;
  556. break;
  557. }
  558. }
  559. if (doCheckUpdate)
  560. {
  561. StringAttr lfn;
  562. Owned<IDistributedFile> file = queryThorFileManager().lookup(queryJob(), filename, temporary, true);
  563. if (file)
  564. {
  565. IPropertyTree &props = file->queryAttributes();
  566. if ((eclCRC == props.getPropInt("@eclCRC")) && (totalCRC == props.getPropInt64("@totalCRC")))
  567. {
  568. // so this needs pruning
  569. Owned<IThorException> e = MakeActivityWarning(this, TE_UpToDate, "output file = '%s' - is up to date - it will not be rebuilt", file->queryLogicalName());
  570. queryOwner().fireException(e);
  571. return true;
  572. }
  573. }
  574. }
  575. return false;
  576. }
  577. void CMasterGraphElement::initActivity()
  578. {
  579. CriticalBlock b(crit);
  580. if (!initialized)
  581. {
  582. initialized = true;
  583. ((CMasterActivity *)queryActivity())->init();
  584. }
  585. }
  586. void CMasterGraphElement::reset()
  587. {
  588. CGraphElementBase::reset();
  589. if (activity && activity->needReInit())
  590. initialized = false;
  591. }
  592. void CMasterGraphElement::doCreateActivity(size32_t parentExtractSz, const byte *parentExtract, MemoryBuffer *startCtx)
  593. {
  594. createActivity();
  595. onStart(parentExtractSz, parentExtract, startCtx);
  596. initActivity();
  597. }
  598. void CMasterGraphElement::slaveDone(size32_t slaveIdx, MemoryBuffer &mb)
  599. {
  600. ((CMasterActivity *)queryActivity())->slaveDone(slaveIdx, mb);
  601. }
  602. //////
  603. ///////
  604. class CBarrierMaster : public CInterface, implements IBarrier
  605. {
  606. mptag_t tag;
  607. Linked<ICommunicator> comm;
  608. bool receiving;
  609. public:
  610. IMPLEMENT_IINTERFACE;
  611. CBarrierMaster(ICommunicator &_comm, mptag_t _tag) : comm(&_comm), tag(_tag)
  612. {
  613. receiving = false;
  614. }
  615. virtual const mptag_t queryTag() const { return tag; }
  616. virtual bool wait(bool exception, unsigned timeout)
  617. {
  618. Owned<IException> e;
  619. CTimeMon tm(timeout);
  620. unsigned s=comm->queryGroup().ordinality()-1;
  621. bool aborted = false;
  622. CMessageBuffer msg;
  623. Owned<IBitSet> raisedSet = createThreadSafeBitSet();
  624. unsigned remaining = timeout;
  625. while (s--)
  626. {
  627. rank_t sender;
  628. msg.clear();
  629. if (INFINITE != timeout && tm.timedout(&remaining))
  630. {
  631. if (exception)
  632. throw createBarrierAbortException();
  633. else
  634. return false;
  635. }
  636. {
  637. BooleanOnOff onOff(receiving);
  638. if (!comm->recv(msg, RANK_ALL, tag, &sender, remaining))
  639. break;
  640. }
  641. msg.read(aborted);
  642. bool hasExcept;
  643. msg.read(hasExcept);
  644. if (hasExcept && !e.get())
  645. e.setown(deserializeException(msg));
  646. sender = sender - 1; // 0 = master
  647. if (raisedSet->testSet(sender, true) && !aborted)
  648. WARNLOG("CBarrierMaster, raise barrier message on tag %d, already received from slave %d", tag, sender);
  649. if (aborted) break;
  650. }
  651. msg.clear();
  652. msg.append(aborted);
  653. if (e)
  654. {
  655. msg.append(true);
  656. serializeException(e, msg);
  657. }
  658. else
  659. msg.append(false);
  660. if (INFINITE != timeout && tm.timedout(&remaining))
  661. {
  662. if (exception)
  663. throw createBarrierAbortException();
  664. else
  665. return false;
  666. }
  667. if (!comm->send(msg, RANK_ALL_OTHER, tag, INFINITE != timeout ? remaining : LONGTIMEOUT))
  668. throw MakeStringException(0, "CBarrierMaster::wait - Timeout sending to slaves");
  669. if (aborted)
  670. {
  671. if (!exception)
  672. return false;
  673. if (e)
  674. throw e.getClear();
  675. else
  676. throw createBarrierAbortException();
  677. }
  678. return true;
  679. }
  680. virtual void cancel(IException *e)
  681. {
  682. if (receiving)
  683. comm->cancel(RANK_ALL, tag);
  684. CMessageBuffer msg;
  685. msg.append(true);
  686. if (e)
  687. {
  688. msg.append(true);
  689. serializeException(e, msg);
  690. }
  691. else
  692. msg.append(false);
  693. if (!comm->send(msg, RANK_ALL_OTHER, tag, LONGTIMEOUT))
  694. throw MakeStringException(0, "CBarrierMaster::cancel - Timeout sending to slaves");
  695. }
  696. };
  697. /////////////
  698. class CMasterGraphTempHandler : public CGraphTempHandler
  699. {
  700. public:
  701. CMasterGraphTempHandler(CJobBase &job, bool errorOnMissing) : CGraphTempHandler(job, errorOnMissing) { }
  702. virtual bool removeTemp(const char *name)
  703. {
  704. queryThorFileManager().clearCacheEntry(name);
  705. return true;
  706. }
  707. virtual void registerFile(const char *name, graph_id graphId, unsigned usageCount, bool temp, WUFileKind fileKind, StringArray *clusters)
  708. {
  709. if (!temp || job.queryUseCheckpoints())
  710. {
  711. Owned<IWorkUnit> wu = &job.queryWorkUnit().lock();
  712. wu->addFile(name, clusters, usageCount, fileKind, job.queryGraphName());
  713. }
  714. else
  715. CGraphTempHandler::registerFile(name, graphId, usageCount, temp, fileKind, clusters);
  716. }
  717. virtual void deregisterFile(const char *name, bool kept) // NB: only called for temp files
  718. {
  719. if (kept || job.queryUseCheckpoints())
  720. {
  721. Owned<IWorkUnit> wu = &job.queryWorkUnit().lock();
  722. wu->releaseFile(name);
  723. }
  724. else
  725. CGraphTempHandler::deregisterFile(name);
  726. }
  727. virtual void clearTemps()
  728. {
  729. try
  730. {
  731. if (!job.queryPausing()) // temps of completed workunit will have been preserved and want to keep
  732. {
  733. Owned<IWorkUnit> lwu = &job.queryWorkUnit().lock();
  734. lwu->deleteTempFiles(job.queryGraphName(), false, false);
  735. }
  736. }
  737. catch (IException *e)
  738. {
  739. EXCLOG(e, "Problem deleting temp files");
  740. e->Release();
  741. }
  742. CGraphTempHandler::clearTemps();
  743. }
  744. };
  745. static const char * getResultText(StringBuffer & s, const char * stepname, unsigned sequence)
  746. {
  747. switch ((int)sequence)
  748. {
  749. case -1: return s.append("STORED('").append(stepname).append("')");
  750. case -2: return s.append("PERSIST('").append(stepname).append("')");
  751. case -3: return s.append("global('").append(stepname).append("')");
  752. default:
  753. if (stepname)
  754. return s.append(stepname);
  755. return s.append('#').append(sequence);
  756. }
  757. }
  758. class CThorCodeContextMaster : public CThorCodeContextBase
  759. {
  760. Linked<IConstWorkUnit> workunit;
  761. Owned<IDistributedFileTransaction> superfiletransaction;
  762. IConstWUResult * getResult(const char * name, unsigned sequence)
  763. {
  764. return getWorkUnitResult(workunit, name, sequence);
  765. }
  766. #define PROTECTED_GETRESULT(STEPNAME, SEQUENCE, KIND, KINDTEXT, ACTION) \
  767. LOG(MCdebugProgress, unknownJob, "getResult%s(%s,%d)", KIND, STEPNAME?STEPNAME:"", SEQUENCE); \
  768. Owned<IConstWUResult> r = getResultForGet(STEPNAME, SEQUENCE); \
  769. try \
  770. { \
  771. ACTION \
  772. } \
  773. catch (IException * e) { \
  774. StringBuffer s, text; e->errorMessage(text); e->Release(); \
  775. throw MakeStringException(TE_FailedToRetrieveWorkunitValue, "result %s in workunit contains an invalid " KINDTEXT " value [%s]", getResultText(s, STEPNAME, SEQUENCE), text.str()); \
  776. } \
  777. catch (CATCHALL) { StringBuffer s; throw MakeStringException(TE_FailedToRetrieveWorkunitValue, "value %s in workunit contains an invalid " KINDTEXT " value", getResultText(s, STEPNAME, SEQUENCE)); }
  778. public:
  779. CThorCodeContextMaster(CJobChannel &jobChannel, IConstWorkUnit &_workunit, ILoadedDllEntry &querySo, IUserDescriptor &userDesc) : CThorCodeContextBase(jobChannel, querySo, userDesc), workunit(&_workunit)
  780. {
  781. }
  782. // ICodeContext
  783. virtual unsigned getGraphLoopCounter() const override { return 0; }
  784. virtual IDebuggableContext *queryDebugContext() const override { return nullptr; }
  785. virtual void setResultBool(const char *name, unsigned sequence, bool result) override
  786. {
  787. WorkunitUpdate w(&workunit->lock());
  788. Owned<IWUResult> r = updateWorkUnitResult(w, name, sequence);
  789. if (r)
  790. {
  791. r->setResultBool(result);
  792. r->setResultStatus(ResultStatusCalculated);
  793. }
  794. else
  795. throw MakeStringException(TE_UnexpectedParameters, "Unexpected parameters to setResultBool");
  796. }
  797. virtual void setResultData(const char *name, unsigned sequence, int len, const void *result) override
  798. {
  799. WorkunitUpdate w(&workunit->lock());
  800. Owned<IWUResult> r = updateWorkUnitResult(w, name, sequence);
  801. if (r)
  802. {
  803. r->setResultData(result, len);
  804. r->setResultStatus(ResultStatusCalculated);
  805. }
  806. else
  807. throw MakeStringException(TE_UnexpectedParameters, "Unexpected parameters to setResultData");
  808. }
  809. virtual void setResultDecimal(const char * name, unsigned sequence, int len, int precision, bool isSigned, const void *val) override
  810. {
  811. WorkunitUpdate w(&workunit->lock());
  812. Owned<IWUResult> r = updateWorkUnitResult(w, name, sequence);
  813. if (r)
  814. {
  815. r->setResultDecimal(val, len);
  816. r->setResultStatus(ResultStatusCalculated);
  817. }
  818. else
  819. throw MakeStringException(TE_UnexpectedParameters, "Unexpected parameters to setResultDecimal");
  820. }
  821. virtual void setResultInt(const char *name, unsigned sequence, __int64 result, unsigned size) override
  822. {
  823. WorkunitUpdate w(&workunit->lock());
  824. Owned<IWUResult> r = updateWorkUnitResult(w, name, sequence);
  825. if (r)
  826. {
  827. r->setResultInt(result);
  828. r->setResultStatus(ResultStatusCalculated);
  829. }
  830. else
  831. throw MakeStringException(TE_UnexpectedParameters, "Unexpected parameters to setResultInt");
  832. }
  833. virtual void setResultRaw(const char *name, unsigned sequence, int len, const void *result) override
  834. {
  835. WorkunitUpdate w(&workunit->lock());
  836. Owned<IWUResult> r = updateWorkUnitResult(w, name, sequence);
  837. if (r)
  838. {
  839. r->setResultRaw(len, result, ResultFormatRaw);
  840. r->setResultStatus(ResultStatusCalculated);
  841. }
  842. else
  843. throw MakeStringException(TE_UnexpectedParameters, "Unexpected parameters to setResultData");
  844. }
  845. virtual void setResultReal(const char *name, unsigned sequence, double result) override
  846. {
  847. WorkunitUpdate w(&workunit->lock());
  848. Owned<IWUResult> r = updateWorkUnitResult(w, name, sequence);
  849. if (r)
  850. {
  851. r->setResultReal(result);
  852. r->setResultStatus(ResultStatusCalculated);
  853. }
  854. else
  855. throw MakeStringException(TE_UnexpectedParameters, "Unexpected parameters to setResultReal");
  856. }
  857. virtual void setResultSet(const char *name, unsigned sequence, bool isAll, size32_t len, const void *result, ISetToXmlTransformer *) override
  858. {
  859. WorkunitUpdate w(&workunit->lock());
  860. Owned<IWUResult> r = updateWorkUnitResult(w, name, sequence);
  861. if (r)
  862. {
  863. r->setResultIsAll(isAll);
  864. r->setResultRaw(len, result, ResultFormatRaw);
  865. r->setResultStatus(ResultStatusCalculated);
  866. }
  867. else
  868. throw MakeStringException(TE_UnexpectedParameters, "Unexpected parameters to setResultData");
  869. }
  870. virtual void setResultString(const char *name, unsigned sequence, int len, const char *result) override
  871. {
  872. WorkunitUpdate w(&workunit->lock());
  873. Owned<IWUResult> r = updateWorkUnitResult(w, name, sequence);
  874. if (r)
  875. {
  876. r->setResultString(result, len);
  877. r->setResultStatus(ResultStatusCalculated);
  878. }
  879. else
  880. throw MakeStringException(TE_UnexpectedParameters, "Unexpected parameters to setResultString");
  881. }
  882. virtual void setResultUnicode(const char * name, unsigned sequence, int len, UChar const * result) override
  883. {
  884. WorkunitUpdate w(&workunit->lock());
  885. Owned<IWUResult> r = updateWorkUnitResult(w, name, sequence);
  886. if (r)
  887. {
  888. r->setResultUnicode(result, len);
  889. r->setResultStatus(ResultStatusCalculated);
  890. }
  891. else
  892. throw MakeStringException(TE_UnexpectedParameters, "Unexpected parameters to setResultUnicode");
  893. }
  894. virtual void setResultVarString(const char * name, unsigned sequence, const char *result) override
  895. {
  896. WorkunitUpdate w(&workunit->lock());
  897. Owned<IWUResult> r = updateWorkUnitResult(w, name, sequence);
  898. if (r)
  899. {
  900. r->setResultString(result, strlen(result));
  901. r->setResultStatus(ResultStatusCalculated);
  902. }
  903. else
  904. throw MakeStringException(TE_UnexpectedParameters, "Unexpected parameters to setResultVarString");
  905. }
  906. virtual void setResultUInt(const char *name, unsigned sequence, unsigned __int64 result, unsigned size) override
  907. {
  908. WorkunitUpdate w(&workunit->lock());
  909. Owned<IWUResult> r = updateWorkUnitResult(w, name, sequence);
  910. if (r)
  911. {
  912. r->setResultUInt(result);
  913. r->setResultStatus(ResultStatusCalculated);
  914. }
  915. else
  916. throw MakeStringException(TE_UnexpectedParameters, "Unexpected parameters to setResultUInt");
  917. }
  918. virtual void setResultVarUnicode(const char * stepname, unsigned sequence, UChar const *val) override
  919. {
  920. setResultUnicode(stepname, sequence, rtlUnicodeStrlen(val), val);
  921. }
  922. virtual bool getResultBool(const char * stepname, unsigned sequence) override
  923. {
  924. PROTECTED_GETRESULT(stepname, sequence, "Bool", "bool",
  925. return r->getResultBool();
  926. );
  927. }
  928. virtual void getResultData(unsigned & tlen, void * & tgt, const char * stepname, unsigned sequence) override
  929. {
  930. PROTECTED_GETRESULT(stepname, sequence, "Data", "data",
  931. SCMStringBuffer result;
  932. r->getResultString(result, false);
  933. tlen = result.length();
  934. tgt = (char *)result.s.detach();
  935. );
  936. }
  937. virtual void getResultDecimal(unsigned tlen, int precision, bool isSigned, void * tgt, const char * stepname, unsigned sequence) override
  938. {
  939. PROTECTED_GETRESULT(stepname, sequence, "Decimal", "decimal",
  940. r->getResultDecimal(tgt, tlen, precision, isSigned);
  941. );
  942. }
  943. virtual void getResultRaw(unsigned & tlen, void * & tgt, const char * stepname, unsigned sequence, IXmlToRowTransformer * xmlTransformer, ICsvToRowTransformer * csvTransformer) override
  944. {
  945. tgt = NULL;
  946. PROTECTED_GETRESULT(stepname, sequence, "Raw", "raw",
  947. Variable2IDataVal result(&tlen, &tgt);
  948. Owned<IXmlToRawTransformer> rawXmlTransformer = createXmlRawTransformer(xmlTransformer);
  949. Owned<ICsvToRawTransformer> rawCsvTransformer = createCsvRawTransformer(csvTransformer);
  950. r->getResultRaw(result, rawXmlTransformer, rawCsvTransformer);
  951. );
  952. }
  953. virtual void getResultSet(bool & isAll, unsigned & tlen, void * & tgt, const char * stepname, unsigned sequence, IXmlToRowTransformer * xmlTransformer, ICsvToRowTransformer * csvTransformer) override
  954. {
  955. tgt = NULL;
  956. PROTECTED_GETRESULT(stepname, sequence, "Raw", "raw",
  957. Variable2IDataVal result(&tlen, &tgt);
  958. Owned<IXmlToRawTransformer> rawXmlTransformer = createXmlRawTransformer(xmlTransformer);
  959. Owned<ICsvToRawTransformer> rawCsvTransformer = createCsvRawTransformer(csvTransformer);
  960. isAll = r->getResultIsAll();
  961. r->getResultRaw(result, rawXmlTransformer, rawCsvTransformer);
  962. );
  963. }
  964. virtual __int64 getResultInt(const char * name, unsigned sequence) override
  965. {
  966. PROTECTED_GETRESULT(name, sequence, "Int", "integer",
  967. return r->getResultInt();
  968. );
  969. }
  970. virtual double getResultReal(const char * name, unsigned sequence) override
  971. {
  972. PROTECTED_GETRESULT(name, sequence, "Real", "real",
  973. return r->getResultReal();
  974. );
  975. }
  976. virtual void getResultString(unsigned & tlen, char * & tgt, const char * stepname, unsigned sequence) override
  977. {
  978. PROTECTED_GETRESULT(stepname, sequence, "String", "string",
  979. SCMStringBuffer result;
  980. r->getResultString(result, false);
  981. tlen = result.length();
  982. tgt = (char *)result.s.detach();
  983. );
  984. }
  985. virtual void getResultStringF(unsigned tlen, char * tgt, const char * stepname, unsigned sequence) override
  986. {
  987. PROTECTED_GETRESULT(stepname, sequence, "String", "string",
  988. SCMStringBuffer result;
  989. r->getResultString(result, false);
  990. rtlStrToStr(tlen, tgt, result.length(), result.s.str());
  991. );
  992. }
  993. virtual void getResultUnicode(unsigned & tlen, UChar * & tgt, const char * stepname, unsigned sequence) override
  994. {
  995. PROTECTED_GETRESULT(stepname, sequence, "Unicode", "unicode",
  996. MemoryBuffer result;
  997. r->getResultUnicode(MemoryBuffer2IDataVal(result));
  998. tlen = result.length()/2;
  999. tgt = (UChar *)malloc(tlen*2);
  1000. memcpy(tgt, result.toByteArray(), tlen*2);
  1001. );
  1002. }
  1003. virtual char * getResultVarString(const char * stepname, unsigned sequence) override
  1004. {
  1005. PROTECTED_GETRESULT(stepname, sequence, "VarString", "string",
  1006. SCMStringBuffer result;
  1007. r->getResultString(result, false);
  1008. return result.s.detach();
  1009. );
  1010. }
  1011. virtual UChar * getResultVarUnicode(const char * stepname, unsigned sequence) override
  1012. {
  1013. PROTECTED_GETRESULT(stepname, sequence, "VarUnicode", "unicode",
  1014. MemoryBuffer result;
  1015. r->getResultUnicode(MemoryBuffer2IDataVal(result));
  1016. unsigned tlen = result.length()/2;
  1017. result.append((UChar)0);
  1018. return (UChar *)result.detach();
  1019. );
  1020. }
  1021. virtual unsigned getResultHash(const char * name, unsigned sequence) override
  1022. {
  1023. PROTECTED_GETRESULT(name, sequence, "Hash", "hash",
  1024. return r->getResultHash();
  1025. );
  1026. }
  1027. virtual unsigned getExternalResultHash(const char * wuid, const char * stepname, unsigned sequence) override
  1028. {
  1029. try
  1030. {
  1031. LOG(MCdebugProgress, unknownJob, "getExternalResultRaw %s", stepname);
  1032. Owned<IConstWUResult> r = getExternalResult(wuid, stepname, sequence);
  1033. return r->getResultHash();
  1034. }
  1035. catch (CATCHALL)
  1036. {
  1037. throw MakeStringException(TE_FailedToRetrieveWorkunitValue, "Failed to retrieve external data hash %s from workunit %s", stepname, wuid);
  1038. }
  1039. }
  1040. virtual void getResultRowset(size32_t & tcount, const byte * * & tgt, const char * stepname, unsigned sequence, IEngineRowAllocator * _rowAllocator, bool isGrouped, IXmlToRowTransformer * xmlTransformer, ICsvToRowTransformer * csvTransformer) override
  1041. {
  1042. tgt = NULL;
  1043. PROTECTED_GETRESULT(stepname, sequence, "Rowset", "rowset",
  1044. MemoryBuffer datasetBuffer;
  1045. MemoryBuffer2IDataVal result(datasetBuffer);
  1046. Owned<IXmlToRawTransformer> rawXmlTransformer = createXmlRawTransformer(xmlTransformer);
  1047. Owned<ICsvToRawTransformer> rawCsvTransformer = createCsvRawTransformer(csvTransformer);
  1048. r->getResultRaw(result, rawXmlTransformer, rawCsvTransformer);
  1049. Owned<IOutputRowDeserializer> deserializer = _rowAllocator->createDiskDeserializer(this);
  1050. rtlDataset2RowsetX(tcount, tgt, _rowAllocator, deserializer, datasetBuffer.length(), datasetBuffer.toByteArray(), isGrouped);
  1051. );
  1052. }
  1053. virtual void getResultDictionary(size32_t & tcount, const byte * * & tgt, IEngineRowAllocator * _rowAllocator, const char * stepname, unsigned sequence, IXmlToRowTransformer * xmlTransformer, ICsvToRowTransformer * csvTransformer, IHThorHashLookupInfo * hasher) override
  1054. {
  1055. tcount = 0;
  1056. tgt = NULL;
  1057. PROTECTED_GETRESULT(stepname, sequence, "Dictionary", "dictionary",
  1058. MemoryBuffer datasetBuffer;
  1059. MemoryBuffer2IDataVal result(datasetBuffer);
  1060. Owned<IXmlToRawTransformer> rawXmlTransformer = createXmlRawTransformer(xmlTransformer);
  1061. Owned<ICsvToRawTransformer> rawCsvTransformer = createCsvRawTransformer(csvTransformer);
  1062. r->getResultRaw(result, rawXmlTransformer, rawCsvTransformer);
  1063. Owned<IOutputRowDeserializer> deserializer = _rowAllocator->createDiskDeserializer(this);
  1064. rtlDeserializeDictionary(tcount, tgt, _rowAllocator, deserializer, datasetBuffer.length(), datasetBuffer.toByteArray());
  1065. );
  1066. }
  1067. virtual void getExternalResultRaw(unsigned & tlen, void * & tgt, const char * wuid, const char * stepname, unsigned sequence, IXmlToRowTransformer * xmlTransformer, ICsvToRowTransformer * csvTransformer) override
  1068. {
  1069. tgt = NULL;
  1070. try
  1071. {
  1072. LOG(MCdebugProgress, unknownJob, "getExternalResultRaw %s", stepname);
  1073. Variable2IDataVal result(&tlen, &tgt);
  1074. Owned<IConstWUResult> r = getExternalResult(wuid, stepname, sequence);
  1075. Owned<IXmlToRawTransformer> rawXmlTransformer = createXmlRawTransformer(xmlTransformer);
  1076. Owned<ICsvToRawTransformer> rawCsvTransformer = createCsvRawTransformer(csvTransformer);
  1077. r->getResultRaw(result, rawXmlTransformer, rawCsvTransformer);
  1078. }
  1079. catch (CATCHALL)
  1080. {
  1081. throw MakeStringException(TE_FailedToRetrieveWorkunitValue, "Failed to retrieve external data value %s from workunit %s", stepname, wuid);
  1082. }
  1083. }
  1084. virtual void addWuException(const char * text, unsigned code, unsigned severity, const char * source) override
  1085. {
  1086. DBGLOG("%s", text);
  1087. try
  1088. {
  1089. Owned<IWorkUnit> w = updateWorkUnit();
  1090. Owned<IWUException> we = w->createException();
  1091. we->setSeverity((ErrorSeverity)severity);
  1092. we->setExceptionMessage(text);
  1093. we->setExceptionSource(source);
  1094. if (code)
  1095. we->setExceptionCode(code);
  1096. }
  1097. catch (IException *E)
  1098. {
  1099. StringBuffer m;
  1100. E->errorMessage(m);
  1101. DBGLOG("Unable to record exception in workunit: %s", m.str());
  1102. E->Release();
  1103. }
  1104. catch (...)
  1105. {
  1106. DBGLOG("Unable to record exception in workunit: unknown exception");
  1107. }
  1108. }
  1109. virtual void addWuAssertFailure(unsigned code, const char * text, const char * filename, unsigned lineno, unsigned column, bool isAbort) override
  1110. {
  1111. DBGLOG("%s", text);
  1112. try
  1113. {
  1114. Owned<IWorkUnit> w = updateWorkUnit();
  1115. unsigned activity = 0;
  1116. addExceptionToWorkunit(w, SeverityError, "user", code, text, filename, lineno, column, activity);
  1117. }
  1118. catch (IException *E)
  1119. {
  1120. StringBuffer m;
  1121. E->errorMessage(m);
  1122. DBGLOG("Unable to record exception in workunit: %s", m.str());
  1123. E->Release();
  1124. }
  1125. catch (...)
  1126. {
  1127. DBGLOG("Unable to record exception in workunit: unknown exception");
  1128. }
  1129. if (isAbort)
  1130. rtlFailOnAssert(); // minimal implementation
  1131. }
  1132. virtual unsigned __int64 getFileOffset(const char *logicalName) override { assertex(false); return 0; }
  1133. virtual unsigned getNodes() override { return jobChannel.queryJob().querySlaves(); }
  1134. virtual unsigned getNodeNum() override { throw MakeThorException(0, "Unsupported. getNodeNum() called in master"); return (unsigned)-1; }
  1135. virtual char *getFilePart(const char *logicalName, bool create=false) override { assertex(false); return NULL; }
  1136. virtual unsigned __int64 getDatasetHash(const char * name, unsigned __int64 hash) override
  1137. {
  1138. unsigned checkSum = 0;
  1139. Owned<IDistributedFile> iDfsFile = queryThorFileManager().lookup(jobChannel.queryJob(), name, false, true, false); // NB: do not update accessed
  1140. if (iDfsFile.get())
  1141. {
  1142. if (iDfsFile->getFileCheckSum(checkSum))
  1143. hash ^= checkSum;
  1144. else
  1145. {
  1146. StringBuffer modifiedStr;
  1147. if (iDfsFile->queryAttributes().getProp("@modified", modifiedStr))
  1148. hash = rtlHash64Data(modifiedStr.length(), modifiedStr.str(), hash);
  1149. // JCS->GH - what's the best thing to do here, if [for some reason] neither are available..
  1150. }
  1151. }
  1152. return hash;
  1153. }
  1154. virtual IDistributedFileTransaction *querySuperFileTransaction() override
  1155. {
  1156. if (!superfiletransaction.get())
  1157. superfiletransaction.setown(createDistributedFileTransaction(userDesc, this));
  1158. return superfiletransaction.get();
  1159. }
  1160. virtual char *getJobName() override
  1161. {
  1162. return strdup(workunit->queryJobName());
  1163. }
  1164. virtual char *getClusterName() override
  1165. {
  1166. return strdup(workunit->queryClusterName());
  1167. }
  1168. virtual char *getGroupName() override
  1169. {
  1170. StringBuffer out;
  1171. if (globals)
  1172. globals->getProp("@nodeGroup",out);
  1173. return out.detach();
  1174. }
  1175. // ICodeContextExt impl.
  1176. virtual IConstWUResult *getExternalResult(const char * wuid, const char *name, unsigned sequence) override
  1177. {
  1178. Owned<IWorkUnitFactory> factory = getWorkUnitFactory();
  1179. Owned<IConstWorkUnit> externalWU = factory->openWorkUnit(wuid);
  1180. if (!externalWU)
  1181. throw MakeStringException(TE_FailedToRetrieveWorkunitValue, "workunit %s not found, retrieving value %s", wuid, name);
  1182. externalWU->remoteCheckAccess(userDesc, false);
  1183. return getWorkUnitResult(externalWU, name, sequence);
  1184. }
  1185. virtual IConstWUResult *getResultForGet(const char *name, unsigned sequence) override
  1186. {
  1187. Owned<IConstWUResult> r = getResult(name, sequence);
  1188. if (!r || (r->getResultStatus() == ResultStatusUndefined))
  1189. {
  1190. StringBuffer s;
  1191. throw MakeStringException(TE_FailedToRetrieveWorkunitValue, "value %s in workunit is undefined", getResultText(s,name,sequence));
  1192. }
  1193. return r.getClear();
  1194. }
  1195. virtual IWorkUnit *updateWorkUnit() const override
  1196. {
  1197. return &workunit->lock();
  1198. }
  1199. };
  1200. class CThorCodeContextMasterSharedMem : public CThorCodeContextMaster
  1201. {
  1202. IThorAllocator *sharedAllocator;
  1203. public:
  1204. CThorCodeContextMasterSharedMem(CJobChannel &jobChannel, IThorAllocator *_sharedAllocator, IConstWorkUnit &_workunit, ILoadedDllEntry &querySo, IUserDescriptor &userDesc)
  1205. : CThorCodeContextMaster(jobChannel, _workunit, querySo, userDesc)
  1206. {
  1207. sharedAllocator = _sharedAllocator;
  1208. }
  1209. virtual IEngineRowAllocator *getRowAllocator(IOutputMetaData * meta, unsigned activityId) const
  1210. {
  1211. return sharedAllocator->getRowAllocator(meta, activityId);
  1212. }
  1213. };
  1214. /////////////
  1215. //
  1216. // CJobMaster
  1217. //
  1218. void loadPlugin(SafePluginMap *pluginMap, const char *_path, const char *name)
  1219. {
  1220. StringBuffer path(_path);
  1221. path.append(name);
  1222. pluginMap->addPlugin(path.str(), name);
  1223. }
  1224. CJobMaster::CJobMaster(IConstWorkUnit &_workunit, const char *graphName, ILoadedDllEntry *querySo, bool _sendSo, const SocketEndpoint &_agentEp)
  1225. : CJobBase(querySo, graphName), workunit(&_workunit), sendSo(_sendSo), agentEp(_agentEp)
  1226. {
  1227. SCMStringBuffer _token, _scope;
  1228. workunit->getScope(_scope);
  1229. workunit->getSecurityToken(_token);
  1230. wuid.set(workunit->queryWuid());
  1231. user.set(workunit->queryUser());
  1232. token.append(_token.str());
  1233. scope.append(_scope.str());
  1234. globalMemoryMB = globals->getPropInt("@masterMemorySize", globals->getPropInt("@globalMemorySize")); // in MB
  1235. numChannels = 1;
  1236. init();
  1237. if (workunit->hasDebugValue("GlobalId"))
  1238. {
  1239. SCMStringBuffer txId;
  1240. workunit->getDebugValue("GlobalId", txId);
  1241. if (txId.length())
  1242. {
  1243. SocketEndpoint thorEp;
  1244. thorEp.setLocalHost(getMachinePortBase());
  1245. logctx->setGlobalId(txId.str(), thorEp, 0);
  1246. VStringBuffer msg("GlobalId: %s", txId.str());
  1247. workunit->getDebugValue("CallerId", txId);
  1248. if (txId.length())
  1249. msg.append(", CallerId: ").append(txId.str());
  1250. txId.set(logctx->queryLocalId());
  1251. if (txId.length())
  1252. msg.append(", LocalId: ").append(txId.str());
  1253. logctx->CTXLOG("%s", msg.str());
  1254. }
  1255. }
  1256. resumed = WUActionResume == workunit->getAction();
  1257. fatalHandler.setown(new CFatalHandler(globals->getPropInt("@fatal_timeout", FATAL_TIMEOUT)));
  1258. querySent = spillsSaved = false;
  1259. nodeDiskUsageCached = false;
  1260. StringBuffer pluginsDir;
  1261. globals->getProp("@pluginsPath", pluginsDir);
  1262. if (pluginsDir.length())
  1263. addPathSepChar(pluginsDir);
  1264. Owned<IConstWUPluginIterator> pluginIter = &workunit->getPlugins();
  1265. ForEach(*pluginIter)
  1266. {
  1267. IConstWUPlugin &plugin = pluginIter->query();
  1268. SCMStringBuffer name;
  1269. plugin.getPluginName(name);
  1270. loadPlugin(pluginMap, pluginsDir.str(), name.str());
  1271. }
  1272. sharedAllocator.setown(::createThorAllocator(globalMemoryMB, 0, 1, memorySpillAtPercentage, *logctx, crcChecking, usePackedAllocator));
  1273. Owned<IMPServer> mpServer = getMPServer();
  1274. addChannel(mpServer);
  1275. slavemptag = allocateMPTag();
  1276. slaveMsgHandler = new CSlaveMessageHandler(*this, slavemptag);
  1277. tmpHandler.setown(createTempHandler(true));
  1278. xgmml.set(graphXGMML);
  1279. }
  1280. CJobMaster::~CJobMaster()
  1281. {
  1282. if (slaveMsgHandler)
  1283. delete slaveMsgHandler;
  1284. freeMPTag(slavemptag);
  1285. tmpHandler.clear();
  1286. }
  1287. void CJobMaster::addChannel(IMPServer *mpServer)
  1288. {
  1289. jobChannels.append(*new CJobMasterChannel(*this, mpServer, jobChannels.ordinality()));
  1290. }
  1291. static IException *createBCastException(unsigned slave, const char *errorMsg)
  1292. {
  1293. // think this should always be fatal, could check link down here, or in general and flag as _shutdown.
  1294. StringBuffer msg("General failure communicating to slave");
  1295. if (slave)
  1296. msg.append("(").append(slave).append(") ");
  1297. else
  1298. msg.append("s ");
  1299. Owned<IThorException> e = MakeThorException(0, "%s", msg.append(" [").append(errorMsg).append("]").str());
  1300. e->setAction(tea_shutdown);
  1301. return e.getClear();
  1302. }
  1303. mptag_t CJobMaster::allocateMPTag()
  1304. {
  1305. mptag_t tag = allocateClusterMPTag();
  1306. queryJobChannel(0).queryJobComm().flush(tag);
  1307. PROGLOG("allocateMPTag: tag = %d", (int)tag);
  1308. return tag;
  1309. }
  1310. void CJobMaster::freeMPTag(mptag_t tag)
  1311. {
  1312. if (TAG_NULL != tag)
  1313. {
  1314. freeClusterMPTag(tag);
  1315. PROGLOG("freeMPTag: tag = %d", (int)tag);
  1316. queryJobChannel(0).queryJobComm().flush(tag);
  1317. }
  1318. }
  1319. void CJobMaster::broadcast(ICommunicator &comm, CMessageBuffer &msg, mptag_t mptag, unsigned timeout, const char *errorMsg, CReplyCancelHandler *msgHandler, bool sendOnly)
  1320. {
  1321. unsigned groupSizeExcludingMaster = comm.queryGroup().ordinality() - 1;
  1322. mptag_t replyTag = TAG_NULL;
  1323. if (!sendOnly)
  1324. {
  1325. replyTag = queryJobChannel(0).queryMPServer().createReplyTag();
  1326. msg.setReplyTag(replyTag);
  1327. }
  1328. if (globals->getPropBool("@broadcastSendAsync", true)) // only here in case of problems/debugging.
  1329. {
  1330. class CSendAsyncfor : public CAsyncFor
  1331. {
  1332. CMessageBuffer &msg;
  1333. mptag_t mptag;
  1334. unsigned timeout;
  1335. StringAttr errorMsg;
  1336. ICommunicator &comm;
  1337. public:
  1338. CSendAsyncfor(ICommunicator &_comm, CMessageBuffer &_msg, mptag_t _mptag, unsigned _timeout, const char *_errorMsg)
  1339. : comm(_comm), msg(_msg), mptag(_mptag), timeout(_timeout), errorMsg(_errorMsg)
  1340. {
  1341. }
  1342. void Do(unsigned i)
  1343. {
  1344. if (!comm.send(msg, i+1, mptag, timeout))
  1345. throw createBCastException(i+1, errorMsg);
  1346. }
  1347. } afor(comm, msg, mptag, timeout, errorMsg);
  1348. try
  1349. {
  1350. afor.For(groupSizeExcludingMaster, groupSizeExcludingMaster);
  1351. }
  1352. catch (IException *e)
  1353. {
  1354. EXCLOG(e, "broadcastSendAsync");
  1355. abort(e);
  1356. throw;
  1357. }
  1358. }
  1359. else if (!comm.send(msg, RANK_ALL_OTHER, mptag, timeout))
  1360. {
  1361. Owned<IException> e = createBCastException(0, errorMsg);
  1362. EXCLOG(e, NULL);
  1363. abort(e);
  1364. throw e.getClear();
  1365. }
  1366. if (sendOnly) return;
  1367. unsigned respondents = 0;
  1368. Owned<IBitSet> bitSet = createThreadSafeBitSet();
  1369. for (;;)
  1370. {
  1371. rank_t sender;
  1372. CMessageBuffer msg;
  1373. bool r = msgHandler ? msgHandler->recv(comm, msg, RANK_ALL, replyTag, &sender, LONGTIMEOUT)
  1374. : comm.recv(msg, RANK_ALL, replyTag, &sender, LONGTIMEOUT);
  1375. if (!r)
  1376. {
  1377. StringBuffer tmpStr;
  1378. if (errorMsg)
  1379. tmpStr.append(": ").append(errorMsg).append(" - ");
  1380. tmpStr.append("Timeout receiving from slaves - no reply from: [");
  1381. unsigned s = bitSet->scan(0, false);
  1382. assertex(s<querySlaves()); // must be at least one
  1383. tmpStr.append(s+1);
  1384. for (;;)
  1385. {
  1386. s = bitSet->scan(s+1, false);
  1387. if (s>=querySlaves())
  1388. break;
  1389. tmpStr.append(",").append(s+1);
  1390. }
  1391. tmpStr.append("]");
  1392. Owned<IException> e = MakeThorFatal(NULL, 0, " %s", tmpStr.str());
  1393. EXCLOG(e, NULL);
  1394. throw e.getClear();
  1395. }
  1396. bool error;
  1397. msg.read(error);
  1398. if (error)
  1399. {
  1400. Owned<IThorException> e = deserializeThorException(msg);
  1401. e->setSlave(sender);
  1402. throw e.getClear();
  1403. }
  1404. ++respondents;
  1405. bitSet->set((unsigned)sender-1);
  1406. if (respondents == groupSizeExcludingMaster)
  1407. break;
  1408. }
  1409. }
  1410. void CJobMaster::initNodeDUCache()
  1411. {
  1412. if (!nodeDiskUsageCached)
  1413. {
  1414. nodeDiskUsageCached = true;
  1415. Owned<IPropertyTreeIterator> fileIter = &workunit->getFileIterator();
  1416. ForEach (*fileIter)
  1417. {
  1418. Owned<IDistributedFile> f = queryDistributedFileDirectory().lookup(fileIter->query().queryProp("@name"), userDesc);
  1419. if (f)
  1420. {
  1421. unsigned n = f->numParts();
  1422. for (unsigned i=0;i<n;i++)
  1423. {
  1424. Owned<IDistributedFilePart> part = f->getPart(i);
  1425. offset_t sz = part->getFileSize(false, false);
  1426. if (i>=nodeDiskUsage.ordinality())
  1427. nodeDiskUsage.append(sz);
  1428. else
  1429. {
  1430. sz += nodeDiskUsage.item(i);
  1431. nodeDiskUsage.add(sz, i);
  1432. }
  1433. }
  1434. }
  1435. }
  1436. }
  1437. }
  1438. IPropertyTree *CJobMaster::prepareWorkUnitInfo()
  1439. {
  1440. Owned<IPropertyTree> workUnitInfo = createPTree("workUnitInfo");
  1441. workUnitInfo->setProp("wuid", wuid);
  1442. workUnitInfo->setProp("user", user);
  1443. workUnitInfo->setProp("token", token);
  1444. workUnitInfo->setProp("scope", scope);
  1445. Owned<IConstWUPluginIterator> pluginIter = &queryWorkUnit().getPlugins();
  1446. IPropertyTree *plugins = NULL;
  1447. ForEach(*pluginIter)
  1448. {
  1449. IConstWUPlugin &thisplugin = pluginIter->query();
  1450. if (!plugins)
  1451. plugins = workUnitInfo->addPropTree("plugins", createPTree());
  1452. SCMStringBuffer name;
  1453. thisplugin.getPluginName(name);
  1454. IPropertyTree *plugin = plugins->addPropTree("plugin", createPTree());
  1455. plugin->setProp("@name", name.str());
  1456. }
  1457. IPropertyTree *debug = workUnitInfo->addPropTree("Debug", createPTree(ipt_caseInsensitive));
  1458. SCMStringBuffer debugStr, valueStr;
  1459. Owned<IStringIterator> debugIter = &queryWorkUnit().getDebugValues();
  1460. ForEach (*debugIter)
  1461. {
  1462. debugIter->str(debugStr);
  1463. queryWorkUnit().getDebugValue(debugStr.str(), valueStr);
  1464. debug->setProp(debugStr.str(), valueStr.str());
  1465. }
  1466. return workUnitInfo.getClear();
  1467. }
  1468. void CJobMaster::sendQuery()
  1469. {
  1470. CriticalBlock b(sendQueryCrit);
  1471. if (querySent) return;
  1472. CMessageBuffer tmp;
  1473. tmp.append(slavemptag);
  1474. tmp.append(queryWuid());
  1475. tmp.append(graphName);
  1476. const char *soName = queryDllEntry().queryName();
  1477. PROGLOG("Query dll: %s", soName);
  1478. tmp.append(soName);
  1479. tmp.append(sendSo);
  1480. if (sendSo)
  1481. {
  1482. CTimeMon atimer;
  1483. OwnedIFile iFile = createIFile(soName);
  1484. OwnedIFileIO iFileIO = iFile->open(IFOread);
  1485. size32_t sz = (size32_t)iFileIO->size();
  1486. tmp.append(sz);
  1487. read(iFileIO, 0, sz, tmp);
  1488. PROGLOG("Loading query for serialization to slaves took %d ms", atimer.elapsed());
  1489. }
  1490. Owned<IPropertyTree> deps = createPTree(queryXGMML()->queryName());
  1491. Owned<IPropertyTreeIterator> edgeIter = queryXGMML()->getElements("edge"); // JCSMORE trim to those actually needed
  1492. ForEach (*edgeIter)
  1493. {
  1494. IPropertyTree &edge = edgeIter->query();
  1495. deps->addPropTree("edge", LINK(&edge));
  1496. }
  1497. Owned<IPropertyTree> workUnitInfo = prepareWorkUnitInfo();
  1498. workUnitInfo->serialize(tmp);
  1499. deps->serialize(tmp);
  1500. CMessageBuffer msg;
  1501. msg.append(QueryInit);
  1502. compressToBuffer(msg, tmp.length(), tmp.toByteArray());
  1503. CTimeMon queryToSlavesTimer;
  1504. broadcast(queryNodeComm(), msg, masterSlaveMpTag, LONGTIMEOUT, "sendQuery");
  1505. PROGLOG("Serialization of query init info (%d bytes) to slaves took %d ms", msg.length(), queryToSlavesTimer.elapsed());
  1506. queryJobManager().addCachedSo(soName);
  1507. querySent = true;
  1508. }
  1509. void CJobMaster::jobDone()
  1510. {
  1511. if (!querySent) return;
  1512. CMessageBuffer msg;
  1513. msg.append(QueryDone);
  1514. msg.append(queryKey());
  1515. broadcast(queryNodeComm(), msg, masterSlaveMpTag, LONGTIMEOUT, "jobDone");
  1516. }
  1517. void CJobMaster::saveSpills()
  1518. {
  1519. CriticalBlock b(spillCrit);
  1520. if (spillsSaved)
  1521. return;
  1522. spillsSaved = true;
  1523. PROGLOG("Paused, saving spills..");
  1524. assertex(!queryUseCheckpoints()); // JCSMORE - checkpoints probably need revisiting
  1525. unsigned numSavedSpills = 0;
  1526. // stash away spills ready for resume, make them owned by workunit in event of abort/delete
  1527. Owned<IFileUsageIterator> iter = queryTempHandler()->getIterator();
  1528. ForEach(*iter)
  1529. {
  1530. CFileUsageEntry &entry = iter->query();
  1531. StringAttr tmpName = entry.queryName();
  1532. if (WUGraphComplete == workunit->queryNodeState(queryGraphName(), entry.queryGraphId()))
  1533. {
  1534. IArrayOf<IGroup> groups;
  1535. StringArray clusters;
  1536. fillClusterArray(*this, tmpName, clusters, groups);
  1537. Owned<IFileDescriptor> fileDesc = queryThorFileManager().create(*this, tmpName, clusters, groups, true, TDXtemporary|TDWnoreplicate);
  1538. fileDesc->queryProperties().setPropBool("@pausefile", true); // JCSMORE - mark to keep, may be able to distinguish via other means
  1539. fileDesc->queryProperties().setProp("@kind", "flat");
  1540. IPropertyTree &props = fileDesc->queryProperties();
  1541. props.setPropBool("@owned", true);
  1542. bool blockCompressed=true; // JCSMORE, should come from helper really
  1543. if (blockCompressed)
  1544. props.setPropBool("@blockCompressed", true);
  1545. Owned<IDistributedFile> file = queryDistributedFileDirectory().createNew(fileDesc);
  1546. // NB: This is renaming/moving from temp path
  1547. StringBuffer newName;
  1548. queryThorFileManager().addScope(*this, tmpName, newName, true, true);
  1549. verifyex(file->renamePhysicalPartFiles(newName.str(), NULL, NULL, queryBaseDirectory(grp_unknown)));
  1550. file->attach(newName,userDesc);
  1551. Owned<IWorkUnit> wu = &queryWorkUnit().lock();
  1552. wu->addFile(newName, &clusters, entry.queryUsage(), entry.queryKind(), queryGraphName());
  1553. ++numSavedSpills;
  1554. }
  1555. }
  1556. PROGLOG("Paused, %d spill(s) saved.", numSavedSpills);
  1557. }
  1558. bool CJobMaster::go()
  1559. {
  1560. class CWorkunitPauseHandler : public CInterface, implements IWorkUnitSubscriber
  1561. {
  1562. CJobMaster &job;
  1563. IConstWorkUnit &wu;
  1564. Owned<IWorkUnitWatcher> watcher;
  1565. CriticalSection crit;
  1566. public:
  1567. IMPLEMENT_IINTERFACE;
  1568. CWorkunitPauseHandler(CJobMaster &_job, IConstWorkUnit &_wu) : job(_job), wu(_wu)
  1569. {
  1570. Owned<IWorkUnitFactory> factory = getWorkUnitFactory();
  1571. watcher.setown(factory->getWatcher(this, (WUSubscribeOptions) (SubscribeOptionAction | SubscribeOptionAbort), wu.queryWuid()));
  1572. }
  1573. ~CWorkunitPauseHandler() { stop(); }
  1574. void stop()
  1575. {
  1576. Owned<IWorkUnitWatcher> _watcher;
  1577. {
  1578. CriticalBlock b(crit);
  1579. if (!watcher)
  1580. return;
  1581. _watcher.setown(watcher.getClear());
  1582. }
  1583. _watcher->unsubscribe();
  1584. }
  1585. void notify(WUSubscribeOptions flags)
  1586. {
  1587. CriticalBlock b(crit);
  1588. if (!watcher)
  1589. return;
  1590. if (flags & SubscribeOptionAbort)
  1591. {
  1592. Owned<IWorkUnitFactory> factory = getWorkUnitFactory();
  1593. if (factory->isAborting(wu.queryWuid()))
  1594. {
  1595. LOG(MCwarning, thorJob, "ABORT detected from user");
  1596. Owned <IException> e = MakeThorException(TE_WorkUnitAborting, "User signalled abort");
  1597. job.fireException(e);
  1598. }
  1599. }
  1600. if (flags & SubscribeOptionAction)
  1601. {
  1602. job.markWuDirty();
  1603. bool abort = false;
  1604. bool pause = false;
  1605. wu.forceReload();
  1606. WUAction action = wu.getAction();
  1607. if (action==WUActionPause)
  1608. {
  1609. // pause after current subgraph
  1610. pause = true;
  1611. }
  1612. else if (action==WUActionPauseNow)
  1613. {
  1614. // abort current subgraph
  1615. abort = true;
  1616. pause = true;
  1617. }
  1618. if (pause)
  1619. {
  1620. PROGLOG("Pausing job%s", abort?" [now]":"");
  1621. job.pause(abort);
  1622. }
  1623. }
  1624. }
  1625. } workunitPauseHandler(*this, *workunit);
  1626. class CQueryTimeoutHandler : public CTimeoutTrigger
  1627. {
  1628. CJobMaster &job;
  1629. public:
  1630. CQueryTimeoutHandler(CJobMaster &_job, unsigned timeout) : CTimeoutTrigger(timeout, "QUERY"), job(_job)
  1631. {
  1632. inform(MakeThorException(TE_QueryTimeoutError, "Query took greater than %d seconds", timeout));
  1633. }
  1634. virtual bool action()
  1635. {
  1636. job.fireException(exception);
  1637. return true;
  1638. }
  1639. private:
  1640. graph_id graphId;
  1641. };
  1642. Owned<CTimeoutTrigger> qtHandler;
  1643. int guillotineTimeout = workunit->getDebugValueInt("maxRunTime", 0);
  1644. if (guillotineTimeout > 0)
  1645. qtHandler.setown(new CQueryTimeoutHandler(*this, guillotineTimeout));
  1646. else if (guillotineTimeout < 0)
  1647. {
  1648. Owned<IException> e = MakeStringException(0, "Ignoring negative maxRunTime: %d", guillotineTimeout);
  1649. reportExceptionToWorkunit(*workunit, e);
  1650. }
  1651. if (WUActionPause == workunit->getAction() || WUActionPauseNow == workunit->getAction())
  1652. throw MakeStringException(0, "Job paused at start, exiting");
  1653. bool allDone = true;
  1654. unsigned concurrentSubGraphs = (unsigned)getWorkUnitValueInt("concurrentSubGraphs", globals->getPropInt("@concurrentSubGraphs", 1));
  1655. try
  1656. {
  1657. startJob();
  1658. workunit->setGraphState(queryGraphName(), WUGraphRunning);
  1659. Owned<IThorGraphIterator> iter = queryJobChannel(0).getSubGraphs();
  1660. CICopyArrayOf<CMasterGraph> toRun;
  1661. ForEach(*iter)
  1662. {
  1663. CMasterGraph &graph = (CMasterGraph &)iter->query();
  1664. if ((queryResumed() || queryUseCheckpoints()) && WUGraphComplete == workunit->queryNodeState(queryGraphName(), graph.queryGraphId()))
  1665. graph.setCompleteEx();
  1666. else
  1667. toRun.append(graph);
  1668. }
  1669. ForEachItemInRev(g, toRun)
  1670. {
  1671. if (aborted) break;
  1672. CMasterGraph &graph = toRun.item(g);
  1673. if (graph.isSink())
  1674. graph.execute(0, NULL, true, concurrentSubGraphs>1);
  1675. if (queryPausing()) break;
  1676. }
  1677. queryJobChannel(0).wait();
  1678. workunitPauseHandler.stop();
  1679. ForEachItemIn(tr, toRun)
  1680. {
  1681. CMasterGraph &graph = toRun.item(tr);
  1682. if (!graph.isComplete())
  1683. {
  1684. allDone = false;
  1685. break;
  1686. }
  1687. }
  1688. }
  1689. catch (IException *e) { fireException(e); e->Release(); }
  1690. catch (CATCHALL) { Owned<IException> e = MakeThorException(0, "Unknown exception running sub graphs"); fireException(e); }
  1691. workunit->setGraphState(queryGraphName(), aborted?WUGraphFailed:(allDone?WUGraphComplete:(pausing?WUGraphPaused:WUGraphComplete)));
  1692. if (queryPausing())
  1693. saveSpills();
  1694. Owned<IException> jobDoneException;
  1695. try { jobDone(); }
  1696. catch (IException *e)
  1697. {
  1698. EXCLOG(e, NULL);
  1699. jobDoneException.setown(e);
  1700. }
  1701. fatalHandler->clear();
  1702. queryTempHandler()->clearTemps();
  1703. slaveMsgHandler->stop();
  1704. if (jobDoneException.get())
  1705. throw LINK(jobDoneException);
  1706. return allDone;
  1707. }
  1708. void CJobMaster::pause(bool doAbort)
  1709. {
  1710. pausing = true;
  1711. if (doAbort)
  1712. {
  1713. // reply will trigger DAMP_THOR_REPLY_PAUSED to agent, unless all graphs are already complete.
  1714. queryJobManager().replyException(*this, NULL);
  1715. // abort current graph asynchronously.
  1716. // After spill files have been saved, trigger timeout handler in case abort doesn't succeed.
  1717. Owned<IException> e = MakeThorException(0, "Unable to recover from pausenow");
  1718. class CAbortThread : implements IThreaded
  1719. {
  1720. CJobMaster &owner;
  1721. CThreaded threaded;
  1722. Linked<IException> exception;
  1723. public:
  1724. CAbortThread(CJobMaster &_owner, IException *_exception) : owner(_owner), exception(_exception), threaded("SaveSpillThread", this)
  1725. {
  1726. threaded.start();
  1727. }
  1728. ~CAbortThread()
  1729. {
  1730. threaded.join();
  1731. }
  1732. // IThreaded
  1733. virtual void threadmain() override
  1734. {
  1735. owner.abort(exception);
  1736. }
  1737. } abortThread(*this, e);
  1738. saveSpills();
  1739. fatalHandler->inform(e.getClear());
  1740. }
  1741. }
  1742. __int64 CJobMaster::queryNodeDiskUsage(unsigned node)
  1743. {
  1744. initNodeDUCache();
  1745. if (!nodeDiskUsage.isItem(node)) return 0;
  1746. return nodeDiskUsage.item(node);
  1747. }
  1748. void CJobMaster::setNodeDiskUsage(unsigned node, __int64 sz)
  1749. {
  1750. initNodeDUCache();
  1751. while (nodeDiskUsage.ordinality() <= node)
  1752. nodeDiskUsage.append(0);
  1753. nodeDiskUsage.replace(sz, node);
  1754. }
  1755. __int64 CJobMaster::addNodeDiskUsage(unsigned node, __int64 sz)
  1756. {
  1757. sz += queryNodeDiskUsage(node);
  1758. setNodeDiskUsage(node, sz);
  1759. return sz;
  1760. }
  1761. bool CJobMaster::queryCreatedFile(const char *file)
  1762. {
  1763. StringBuffer scopedName;
  1764. queryThorFileManager().addScope(*this, file, scopedName, false);
  1765. return (NotFound != createdFiles.find(scopedName.str()));
  1766. }
  1767. void CJobMaster::addCreatedFile(const char *file)
  1768. {
  1769. StringBuffer scopedName;
  1770. queryThorFileManager().addScope(*this, file, scopedName, false);
  1771. createdFiles.append(scopedName.str());
  1772. }
  1773. __int64 CJobMaster::getWorkUnitValueInt(const char *prop, __int64 defVal) const
  1774. {
  1775. return queryWorkUnit().getDebugValueInt64(prop, defVal);
  1776. }
  1777. bool CJobMaster::getWorkUnitValueBool(const char *prop, bool defVal) const
  1778. {
  1779. return queryWorkUnit().getDebugValueBool(prop, defVal);
  1780. }
  1781. StringBuffer &CJobMaster::getWorkUnitValue(const char *prop, StringBuffer &str) const
  1782. {
  1783. SCMStringBuffer scmStr;
  1784. queryWorkUnit().getDebugValue(prop, scmStr);
  1785. return str.append(scmStr.str());
  1786. }
  1787. IGraphTempHandler *CJobMaster::createTempHandler(bool errorOnMissing)
  1788. {
  1789. return new CMasterGraphTempHandler(*this, errorOnMissing);
  1790. }
  1791. bool CJobMaster::fireException(IException *e)
  1792. {
  1793. IThorException *te = QUERYINTERFACE(e, IThorException);
  1794. ThorExceptionAction action;
  1795. if (!te) action = tea_null;
  1796. else action = te->queryAction();
  1797. if (QUERYINTERFACE(e, IMP_Exception) && MPERR_link_closed==e->errorCode())
  1798. action = tea_shutdown;
  1799. else if (QUERYINTERFACE(e, ISEH_Exception))
  1800. action = tea_shutdown;
  1801. CriticalBlock b(exceptCrit);
  1802. switch (action)
  1803. {
  1804. case tea_warning:
  1805. {
  1806. LOG(MCwarning, thorJob, e);
  1807. ErrorSeverity mappedSeverity = workunit->getWarningSeverity(e->errorCode(), SeverityWarning);
  1808. if (mappedSeverity != SeverityIgnore)
  1809. reportExceptionToWorkunit(*workunit, e);
  1810. break;
  1811. }
  1812. default:
  1813. {
  1814. LOG(MCerror, thorJob, e);
  1815. queryJobManager().replyException(*this, e);
  1816. fatalHandler->inform(LINK(e));
  1817. try { abort(e); }
  1818. catch (IException *e)
  1819. {
  1820. Owned<IThorException> te = ThorWrapException(e, "Error aborting job, will cause thor restart");
  1821. e->Release();
  1822. reportExceptionToWorkunit(*workunit, te);
  1823. action = tea_shutdown;
  1824. }
  1825. if (tea_shutdown == action)
  1826. queryJobManager().stop();
  1827. }
  1828. }
  1829. return true;
  1830. }
  1831. // CJobMasterChannel
  1832. CJobMasterChannel::CJobMasterChannel(CJobBase &job, IMPServer *mpServer, unsigned channel) : CJobChannel(job, mpServer, channel)
  1833. {
  1834. codeCtx.setown(new CThorCodeContextMaster(*this, job.queryWorkUnit(), job.queryDllEntry(), *job.queryUserDescriptor()));
  1835. sharedMemCodeCtx.setown(new CThorCodeContextMasterSharedMem(*this, job.querySharedAllocator(), job.queryWorkUnit(), job.queryDllEntry(), *job.queryUserDescriptor()));
  1836. }
  1837. CGraphBase *CJobMasterChannel::createGraph()
  1838. {
  1839. return new CMasterGraph(*this);
  1840. }
  1841. IBarrier *CJobMasterChannel::createBarrier(mptag_t tag)
  1842. {
  1843. return new CBarrierMaster(*jobComm, tag);
  1844. }
  1845. ///////////////////
  1846. class CCollatedResult : implements IThorResult, public CSimpleInterface
  1847. {
  1848. CMasterGraph &graph;
  1849. CActivityBase &activity;
  1850. IThorRowInterfaces *rowIf;
  1851. unsigned id;
  1852. CriticalSection crit;
  1853. PointerArrayOf<CThorExpandingRowArray> results;
  1854. Owned<IThorResult> result;
  1855. unsigned spillPriority;
  1856. activity_id ownerId;
  1857. void ensure()
  1858. {
  1859. CriticalBlock b(crit);
  1860. if (result)
  1861. return;
  1862. mptag_t replyTag = graph.queryMPServer().createReplyTag();
  1863. CMessageBuffer msg;
  1864. msg.append(GraphGetResult);
  1865. msg.append(activity.queryJob().queryKey());
  1866. msg.append(graph.queryGraphId());
  1867. msg.append(ownerId);
  1868. msg.append(id);
  1869. msg.append(replyTag);
  1870. ((CJobMaster &)graph.queryJob()).broadcast(queryNodeComm(), msg, masterSlaveMpTag, LONGTIMEOUT, "CCollectResult", NULL, true);
  1871. unsigned numSlaves = graph.queryJob().querySlaves();
  1872. for (unsigned n=0; n<numSlaves; n++)
  1873. results.item(n)->kill();
  1874. rank_t sender;
  1875. MemoryBuffer mb;
  1876. Owned<ISerialStream> stream = createMemoryBufferSerialStream(mb);
  1877. CThorStreamDeserializerSource rowSource(stream);
  1878. unsigned todo = numSlaves;
  1879. for (;;)
  1880. {
  1881. for (;;)
  1882. {
  1883. if (activity.queryAbortSoon())
  1884. return;
  1885. msg.clear();
  1886. if (activity.receiveMsg(msg, RANK_ALL, replyTag, &sender, 60*1000))
  1887. break;
  1888. ActPrintLog(&activity, "WARNING: tag %d timedout, retrying", (unsigned)replyTag);
  1889. }
  1890. sender = sender - 1; // 0 = master
  1891. if (!msg.length())
  1892. {
  1893. --todo;
  1894. if (0 == todo)
  1895. break; // done
  1896. }
  1897. else
  1898. {
  1899. bool error;
  1900. msg.read(error);
  1901. if (error)
  1902. {
  1903. Owned<IThorException> e = deserializeThorException(msg);
  1904. e->setSlave(sender);
  1905. throw e.getClear();
  1906. }
  1907. ThorExpand(msg, mb.clear());
  1908. CThorExpandingRowArray *slaveResults = results.item(sender);
  1909. while (!rowSource.eos())
  1910. {
  1911. RtlDynamicRowBuilder rowBuilder(rowIf->queryRowAllocator());
  1912. size32_t sz = rowIf->queryRowDeserializer()->deserialize(rowBuilder, rowSource);
  1913. slaveResults->append(rowBuilder.finalizeRowClear(sz));
  1914. }
  1915. }
  1916. }
  1917. Owned<IThorResult> _result = ::createResult(activity, rowIf, thorgraphresult_nul, spillPriority);
  1918. Owned<IRowWriter> resultWriter = _result->getWriter();
  1919. for (unsigned s=0; s<numSlaves; s++)
  1920. {
  1921. CThorExpandingRowArray *slaveResult = results.item(s);
  1922. ForEachItemIn(r, *slaveResult)
  1923. {
  1924. const void *row = slaveResult->query(r);
  1925. LinkThorRow(row);
  1926. resultWriter->putRow(row);
  1927. }
  1928. }
  1929. result.setown(_result.getClear());
  1930. }
  1931. public:
  1932. IMPLEMENT_IINTERFACE_USING(CSimpleInterface);
  1933. CCollatedResult(CMasterGraph &_graph, CActivityBase &_activity, IThorRowInterfaces *_rowIf, unsigned _id, activity_id _ownerId, unsigned _spillPriority)
  1934. : graph(_graph), activity(_activity), rowIf(_rowIf), id(_id), ownerId(_ownerId), spillPriority(_spillPriority)
  1935. {
  1936. for (unsigned n=0; n<graph.queryJob().querySlaves(); n++)
  1937. results.append(new CThorExpandingRowArray(activity, rowIf));
  1938. }
  1939. ~CCollatedResult()
  1940. {
  1941. ForEachItemIn(l, results)
  1942. {
  1943. CThorExpandingRowArray *result = results.item(l);
  1944. delete result;
  1945. }
  1946. }
  1947. void setId(unsigned _id)
  1948. {
  1949. id = _id;
  1950. }
  1951. // IThorResult
  1952. virtual IRowWriter *getWriter() { throwUnexpected(); }
  1953. virtual void setResultStream(IRowWriterMultiReader *stream, rowcount_t count)
  1954. {
  1955. throwUnexpected();
  1956. }
  1957. virtual IRowStream *getRowStream()
  1958. {
  1959. ensure();
  1960. return result->getRowStream();
  1961. }
  1962. virtual IThorRowInterfaces *queryRowInterfaces()
  1963. {
  1964. return rowIf;
  1965. }
  1966. virtual CActivityBase *queryActivity()
  1967. {
  1968. return &activity;
  1969. }
  1970. virtual bool isDistributed() const { return false; }
  1971. virtual void serialize(MemoryBuffer &mb)
  1972. {
  1973. ensure();
  1974. result->serialize(mb);
  1975. }
  1976. virtual void getLinkedResult(unsigned & count, const byte * * & ret) override
  1977. {
  1978. ensure();
  1979. result->getLinkedResult(count, ret);
  1980. }
  1981. virtual const void * getLinkedRowResult()
  1982. {
  1983. ensure();
  1984. return result->getLinkedRowResult();
  1985. }
  1986. };
  1987. ///////////////////
  1988. //
  1989. // CMasterGraph impl.
  1990. //
  1991. CMasterGraph::CMasterGraph(CJobChannel &jobChannel) : CGraphBase(jobChannel)
  1992. {
  1993. jobM = (CJobMaster *)&jobChannel.queryJob();
  1994. mpTag = queryJob().allocateMPTag();
  1995. startBarrierTag = queryJob().allocateMPTag();
  1996. waitBarrierTag = queryJob().allocateMPTag();
  1997. startBarrier = jobChannel.createBarrier(startBarrierTag);
  1998. waitBarrier = jobChannel.createBarrier(waitBarrierTag);
  1999. }
  2000. CMasterGraph::~CMasterGraph()
  2001. {
  2002. queryJob().freeMPTag(mpTag);
  2003. queryJob().freeMPTag(startBarrierTag);
  2004. queryJob().freeMPTag(waitBarrierTag);
  2005. if (TAG_NULL != doneBarrierTag)
  2006. queryJob().freeMPTag(doneBarrierTag);
  2007. if (TAG_NULL != executeReplyTag)
  2008. queryJob().freeMPTag(executeReplyTag);
  2009. }
  2010. void CMasterGraph::init()
  2011. {
  2012. CGraphBase::init();
  2013. if (queryOwner() && isGlobal())
  2014. {
  2015. doneBarrierTag = queryJob().allocateMPTag();
  2016. doneBarrier = jobChannel.createBarrier(doneBarrierTag);
  2017. }
  2018. }
  2019. bool CMasterGraph::fireException(IException *e)
  2020. {
  2021. IThorException *te = QUERYINTERFACE(e, IThorException);
  2022. ThorExceptionAction action;
  2023. if (!te) action = tea_null;
  2024. else action = te->queryAction();
  2025. if (QUERYINTERFACE(e, IMP_Exception) && MPERR_link_closed==e->errorCode())
  2026. action = tea_shutdown;
  2027. else if (QUERYINTERFACE(e, ISEH_Exception))
  2028. action = tea_shutdown;
  2029. CriticalBlock b(exceptCrit);
  2030. switch (action)
  2031. {
  2032. case tea_warning:
  2033. {
  2034. LOG(MCwarning, thorJob, e);
  2035. ErrorSeverity mappedSeverity = job.queryWorkUnit().getWarningSeverity(e->errorCode(), SeverityWarning);
  2036. if (mappedSeverity != SeverityIgnore)
  2037. reportExceptionToWorkunit(job.queryWorkUnit(), e);
  2038. break;
  2039. }
  2040. default:
  2041. {
  2042. LOG(MCerror, thorJob, e);
  2043. if (NULL != fatalHandler)
  2044. fatalHandler->inform(LINK(e));
  2045. if (owner)
  2046. owner->fireException(e);
  2047. else
  2048. queryJobChannel().fireException(e);
  2049. }
  2050. }
  2051. return true;
  2052. }
  2053. void CMasterGraph::reset()
  2054. {
  2055. CGraphBase::reset();
  2056. bcastMsgHandler.reset();
  2057. activityInitMsgHandler.reset();
  2058. executeReplyMsgHandler.reset();
  2059. }
  2060. void CMasterGraph::abort(IException *e)
  2061. {
  2062. if (aborted) return;
  2063. bool _graphDone = graphDone; // aborting master activities can trigger master graphDone, but want to fire GraphAbort to slaves if graphDone=false at start.
  2064. try { CGraphBase::abort(e); }
  2065. catch (IException *e)
  2066. {
  2067. GraphPrintLog(e, "Aborting master graph");
  2068. e->Release();
  2069. }
  2070. bcastMsgHandler.cancel(0);
  2071. activityInitMsgHandler.cancel(RANK_ALL);
  2072. executeReplyMsgHandler.cancel(RANK_ALL);
  2073. if (started && !_graphDone)
  2074. {
  2075. try
  2076. {
  2077. CMessageBuffer msg;
  2078. msg.append(GraphAbort);
  2079. msg.append(job.queryKey());
  2080. msg.append(queryGraphId());
  2081. jobM->broadcast(queryNodeComm(), msg, masterSlaveMpTag, LONGTIMEOUT, "abort");
  2082. }
  2083. catch (IException *e)
  2084. {
  2085. GraphPrintLog(e, "Aborting slave graph");
  2086. if (abortException)
  2087. {
  2088. e->Release();
  2089. throw LINK(abortException);
  2090. }
  2091. throw;
  2092. }
  2093. }
  2094. if (!queryOwner())
  2095. {
  2096. if (globals->getPropBool("@watchdogProgressEnabled"))
  2097. queryJobManager().queryDeMonServer()->endGraph(this, !queryAborted());
  2098. }
  2099. }
  2100. bool CMasterGraph::serializeActivityInitData(unsigned slave, MemoryBuffer &mb, IThorActivityIterator &iter)
  2101. {
  2102. CriticalBlock b(createdCrit);
  2103. DelayedSizeMarker sizeMark1(mb);
  2104. ForEach (iter)
  2105. {
  2106. CMasterGraphElement &element = (CMasterGraphElement &)iter.query();
  2107. if (!element.sentActInitData->testSet(slave))
  2108. {
  2109. CMasterActivity *activity = (CMasterActivity *)element.queryActivity();
  2110. if (activity)
  2111. {
  2112. mb.append(element.queryId());
  2113. DelayedSizeMarker sizeMark2(mb);
  2114. activity->serializeSlaveData(mb, slave);
  2115. sizeMark2.write();
  2116. }
  2117. }
  2118. }
  2119. if (0 == sizeMark1.size())
  2120. return false;
  2121. mb.append((activity_id)0); // terminator
  2122. sizeMark1.write();
  2123. return true;
  2124. }
  2125. void CMasterGraph::readActivityInitData(MemoryBuffer &mb, unsigned slave)
  2126. {
  2127. for (;;)
  2128. {
  2129. activity_id id;
  2130. mb.read(id);
  2131. if (0 == id)
  2132. break;
  2133. size32_t dataLen;
  2134. mb.read(dataLen);
  2135. CGraphElementBase *element = queryElement(id);
  2136. MemoryBuffer &mbDst = element->queryActivity()->queryInitializationData(slave);
  2137. mbDst.append(dataLen, mb.readDirect(dataLen));
  2138. }
  2139. }
  2140. bool CMasterGraph::prepare(size32_t parentExtractSz, const byte *parentExtract, bool checkDependencies, bool shortCircuit, bool async)
  2141. {
  2142. if (!CGraphBase::prepare(parentExtractSz, parentExtract, checkDependencies, shortCircuit, async)) return false;
  2143. if (aborted) return false;
  2144. return true;
  2145. }
  2146. void CMasterGraph::execute(size32_t _parentExtractSz, const byte *parentExtract, bool checkDependencies, bool async)
  2147. {
  2148. if (isComplete())
  2149. return;
  2150. if (!queryOwner()) // owning graph sends query+child graphs
  2151. jobM->sendQuery(); // if not previously sent
  2152. CGraphBase::execute(parentExtractSz, parentExtract, checkDependencies, async);
  2153. }
  2154. void CMasterGraph::start()
  2155. {
  2156. if (!queryOwner())
  2157. {
  2158. if (globals->getPropBool("@watchdogProgressEnabled"))
  2159. queryJobManager().queryDeMonServer()->startGraph(this);
  2160. }
  2161. Owned<IThorActivityIterator> iter = getConnectedIterator();
  2162. ForEach (*iter)
  2163. iter->query().queryActivity()->startProcess();
  2164. }
  2165. void CMasterGraph::sendActivityInitData()
  2166. {
  2167. CMessageBuffer msg;
  2168. mptag_t replyTag = queryJobChannel().queryMPServer().createReplyTag();
  2169. msg.setReplyTag(replyTag);
  2170. unsigned pos = msg.length();
  2171. unsigned w=0;
  2172. unsigned sentTo = 0;
  2173. Owned<IThorActivityIterator> iter = getConnectedIterator();
  2174. for (; w<queryJob().querySlaves(); w++)
  2175. {
  2176. unsigned needActInit = 0;
  2177. ForEach(*iter)
  2178. {
  2179. CGraphElementBase &element = iter->query();
  2180. CActivityBase *activity = element.queryActivity();
  2181. if (activity && activity->needReInit())
  2182. element.sentActInitData->set(w, false); // force act init to be resent
  2183. if (!element.sentActInitData->test(w)) // has it been sent
  2184. ++needActInit;
  2185. }
  2186. if (needActInit)
  2187. {
  2188. try
  2189. {
  2190. msg.rewrite(pos);
  2191. serializeActivityInitData(w, msg, *iter);
  2192. }
  2193. catch (IException *e)
  2194. {
  2195. GraphPrintLog(e);
  2196. throw;
  2197. }
  2198. if (!queryJobChannel().queryJobComm().send(msg, w+1, mpTag, LONGTIMEOUT))
  2199. {
  2200. StringBuffer epStr;
  2201. throw MakeStringException(0, "Timeout sending to slave %s", job.querySlaveGroup().queryNode(w).endpoint().getUrlStr(epStr).str());
  2202. }
  2203. ++sentTo;
  2204. }
  2205. }
  2206. if (sentTo)
  2207. {
  2208. assertex(sentTo == queryJob().querySlaves());
  2209. w=0;
  2210. Owned<IException> e;
  2211. // now get back initialization data from graph tag
  2212. for (; w<queryJob().querySlaves(); w++)
  2213. {
  2214. rank_t sender;
  2215. msg.clear();
  2216. if (!activityInitMsgHandler.recv(queryJobChannel().queryJobComm(), msg, w+1, replyTag, &sender, LONGTIMEOUT))
  2217. throw MakeGraphException(this, 0, "Timeout receiving from slaves after graph sent");
  2218. bool error;
  2219. msg.read(error);
  2220. if (error)
  2221. {
  2222. Owned<IThorException> se = deserializeThorException(msg);
  2223. se->setSlave(sender);
  2224. if (!e.get())
  2225. {
  2226. StringBuffer tmpStr("Slave ");
  2227. queryJob().queryJobGroup().queryNode(sender).endpoint().getUrlStr(tmpStr);
  2228. GraphPrintLog(se, "%s", tmpStr.append(": slave initialization error").str());
  2229. e.setown(se.getClear());
  2230. }
  2231. continue; // to read other slave responses.
  2232. }
  2233. readActivityInitData(msg, w);
  2234. }
  2235. if (e.get())
  2236. throw LINK(e);
  2237. }
  2238. }
  2239. void CMasterGraph::serializeGraphInit(MemoryBuffer &mb)
  2240. {
  2241. mb.append(graphId);
  2242. serializeMPtag(mb, mpTag);
  2243. mb.append((int)startBarrierTag);
  2244. mb.append((int)waitBarrierTag);
  2245. mb.append((int)doneBarrierTag);
  2246. mb.append(queryChildGraphCount());
  2247. Owned<IThorGraphIterator> childIter = getChildGraphIterator();
  2248. ForEach (*childIter)
  2249. {
  2250. CMasterGraph &childGraph = (CMasterGraph &)childIter->query();
  2251. childGraph.serializeGraphInit(mb);
  2252. }
  2253. }
  2254. // IThorChildGraph impl.
  2255. IEclGraphResults *CMasterGraph::evaluate(unsigned _parentExtractSz, const byte *parentExtract)
  2256. {
  2257. throw MakeGraphException(this, 0, "Thor master does not support the execution of child queries");
  2258. }
  2259. void CMasterGraph::executeSubGraph(size32_t parentExtractSz, const byte *parentExtract)
  2260. {
  2261. if (job.queryResumed()) // skip complete subgraph if resuming. NB: temp spill have been tucked away for this purpose when paused.
  2262. {
  2263. if (!queryOwner())
  2264. {
  2265. if (WUGraphComplete == job.queryWorkUnit().queryNodeState(job.queryGraphName(), graphId))
  2266. setCompleteEx();
  2267. }
  2268. }
  2269. if (isComplete())
  2270. return;
  2271. if (!sentGlobalInit)
  2272. {
  2273. sentGlobalInit = true;
  2274. if (!queryOwner())
  2275. sendGraph();
  2276. else
  2277. {
  2278. if (isGlobal())
  2279. {
  2280. CMessageBuffer msg;
  2281. serializeCreateContexts(msg);
  2282. try
  2283. {
  2284. jobM->broadcast(queryJobChannel().queryJobComm(), msg, mpTag, LONGTIMEOUT, "serializeCreateContexts", &bcastMsgHandler);
  2285. }
  2286. catch (IException *e)
  2287. {
  2288. GraphPrintLog(e, "Aborting graph create(2)");
  2289. if (abortException)
  2290. {
  2291. e->Release();
  2292. throw LINK(abortException);
  2293. }
  2294. throw;
  2295. }
  2296. }
  2297. }
  2298. }
  2299. fatalHandler.clear();
  2300. fatalHandler.setown(new CFatalHandler(globals->getPropInt("@fatal_timeout", FATAL_TIMEOUT)));
  2301. CGraphBase::executeSubGraph(parentExtractSz, parentExtract);
  2302. if (TAG_NULL != executeReplyTag)
  2303. {
  2304. rank_t sender;
  2305. unsigned s=0;
  2306. for (; s<queryJob().querySlaves(); s++)
  2307. {
  2308. CMessageBuffer msg;
  2309. if (!executeReplyMsgHandler.recv(queryJobChannel().queryJobComm(), msg, RANK_ALL, executeReplyTag, &sender))
  2310. break;
  2311. bool error;
  2312. msg.read(error);
  2313. if (error)
  2314. {
  2315. Owned<IThorException> exception = deserializeThorException(msg);
  2316. exception->setSlave(sender);
  2317. GraphPrintLog(exception, "slave execute reply exception");
  2318. throw exception.getClear();
  2319. }
  2320. }
  2321. if (fatalHandler)
  2322. fatalHandler->clear();
  2323. }
  2324. fatalHandler.clear();
  2325. Owned<IWorkUnit> wu = &job.queryWorkUnit().lock();
  2326. queryJobManager().updateWorkUnitLog(*wu);
  2327. }
  2328. void CMasterGraph::sendGraph()
  2329. {
  2330. CTimeMon atimer;
  2331. CMessageBuffer msg;
  2332. msg.append(GraphInit);
  2333. msg.append(job.queryKey());
  2334. if (TAG_NULL == executeReplyTag)
  2335. executeReplyTag = jobM->allocateMPTag();
  2336. serializeMPtag(msg, executeReplyTag);
  2337. serializeCreateContexts(msg);
  2338. serializeGraphInit(msg);
  2339. // slave graph data
  2340. try
  2341. {
  2342. jobM->broadcast(queryNodeComm(), msg, masterSlaveMpTag, LONGTIMEOUT, "sendGraph", &bcastMsgHandler);
  2343. }
  2344. catch (IException *e)
  2345. {
  2346. GraphPrintLog(e, "Aborting sendGraph");
  2347. if (abortException)
  2348. {
  2349. e->Release();
  2350. throw LINK(abortException);
  2351. }
  2352. throw;
  2353. }
  2354. GraphPrintLog("sendGraph took %d ms", atimer.elapsed());
  2355. }
  2356. bool CMasterGraph::preStart(size32_t parentExtractSz, const byte *parentExtract)
  2357. {
  2358. GraphPrintLog("Processing graph");
  2359. if (syncInitData())
  2360. {
  2361. sendActivityInitData(); // has to be done at least once
  2362. // NB: At this point, on the slaves, the graphs will start
  2363. }
  2364. CGraphBase::preStart(parentExtractSz, parentExtract);
  2365. if (isGlobal())
  2366. {
  2367. if (!startBarrier->wait(true)) // ensure all graphs are at start point at same time, as some may request data from each other.
  2368. return false;
  2369. }
  2370. if (!queryOwner())
  2371. job.queryWorkUnit().setNodeState(job.queryGraphName(), graphId, WUGraphRunning);
  2372. return true;
  2373. }
  2374. void CMasterGraph::handleSlaveDone(unsigned node, MemoryBuffer &mb)
  2375. {
  2376. graph_id gid;
  2377. mb.read(gid);
  2378. assertex(gid == queryGraphId());
  2379. unsigned count;
  2380. mb.read(count);
  2381. while (count--)
  2382. {
  2383. activity_id activityId;
  2384. mb.read(activityId);
  2385. CMasterGraphElement *act = (CMasterGraphElement *)queryElement(activityId);
  2386. unsigned len;
  2387. mb.read(len);
  2388. const void *d = mb.readDirect(len);
  2389. MemoryBuffer sdMb;
  2390. sdMb.setBuffer(len, (void *)d);
  2391. act->slaveDone(node, sdMb);
  2392. }
  2393. }
  2394. void CMasterGraph::getFinalProgress()
  2395. {
  2396. offset_t totalDiskUsage = 0;
  2397. offset_t minNodeDiskUsage = 0, maxNodeDiskUsage = 0;
  2398. unsigned maxNode = (unsigned)-1, minNode = (unsigned)-1;
  2399. CMessageBuffer msg;
  2400. mptag_t replyTag = queryJobChannel().queryMPServer().createReplyTag();
  2401. msg.setReplyTag(replyTag);
  2402. msg.append((unsigned)GraphEnd);
  2403. msg.append(job.queryKey());
  2404. msg.append(queryGraphId());
  2405. jobM->broadcast(queryNodeComm(), msg, masterSlaveMpTag, LONGTIMEOUT, "graphEnd", NULL, true);
  2406. Owned<IBitSet> respondedBitSet = createBitSet();
  2407. unsigned n=queryJob().queryNodes();
  2408. while (n--)
  2409. {
  2410. rank_t sender;
  2411. if (!queryNodeComm().recv(msg, RANK_ALL, replyTag, &sender, LONGTIMEOUT))
  2412. {
  2413. StringBuffer slaveList;
  2414. n=queryJob().queryNodes();
  2415. unsigned s = 0;
  2416. for (;;)
  2417. {
  2418. unsigned ns = respondedBitSet->scan(s, true); // scan for next respondent
  2419. // list all non-respondents up to this found respondent (or rest of slaves)
  2420. while (s<ns && s<n) // NB: if scan find nothing else set, ns = (unsigned)-1
  2421. {
  2422. if (slaveList.length())
  2423. slaveList.append(",");
  2424. ++s; // inc. 1st as slaveList is list of slaves that are 1 based.
  2425. slaveList.append(s);
  2426. }
  2427. ++s;
  2428. if (s>=n)
  2429. break;
  2430. }
  2431. throw MakeGraphException(this, 0, "Timeout receiving final progress from slaves - these slaves failed to respond: %s", slaveList.str());
  2432. }
  2433. bool error;
  2434. msg.read(error);
  2435. if (error)
  2436. {
  2437. Owned<IThorException> e = deserializeThorException(msg);
  2438. e->setSlave(sender);
  2439. throw e.getClear();
  2440. }
  2441. respondedBitSet->set(((unsigned)sender)-1);
  2442. if (0 == msg.remaining())
  2443. continue;
  2444. unsigned channelsPerSlave = globals->getPropInt("@channelsPerSlave", 1); // JCSMORE - should move somewhere common
  2445. for (unsigned sc=0; sc<channelsPerSlave; sc++)
  2446. {
  2447. unsigned slave;
  2448. msg.read(slave);
  2449. handleSlaveDone(slave, msg);
  2450. if (!queryOwner())
  2451. {
  2452. if (globals->getPropBool("@watchdogProgressEnabled"))
  2453. {
  2454. try
  2455. {
  2456. size32_t progressLen;
  2457. msg.read(progressLen);
  2458. MemoryBuffer progressData;
  2459. progressData.setBuffer(progressLen, (void *)msg.readDirect(progressLen));
  2460. queryJobManager().queryDeMonServer()->takeHeartBeat(progressData);
  2461. }
  2462. catch (IException *e)
  2463. {
  2464. GraphPrintLog(e, "Failure whilst deserializing stats/progress");
  2465. e->Release();
  2466. }
  2467. }
  2468. }
  2469. offset_t nodeDiskUsage;
  2470. msg.read(nodeDiskUsage);
  2471. jobM->setNodeDiskUsage(n, nodeDiskUsage);
  2472. if (nodeDiskUsage > maxNodeDiskUsage)
  2473. {
  2474. maxNodeDiskUsage = nodeDiskUsage;
  2475. maxNode = n;
  2476. }
  2477. if ((unsigned)-1 == minNode || nodeDiskUsage < minNodeDiskUsage)
  2478. {
  2479. minNodeDiskUsage = nodeDiskUsage;
  2480. minNode = n;
  2481. }
  2482. totalDiskUsage += nodeDiskUsage;
  2483. }
  2484. }
  2485. if (totalDiskUsage)
  2486. {
  2487. Owned<IWorkUnit> wu = &job.queryWorkUnit().lock();
  2488. wu->addDiskUsageStats(totalDiskUsage/queryJob().querySlaves(), minNode, minNodeDiskUsage, maxNode, maxNodeDiskUsage, queryGraphId());
  2489. }
  2490. }
  2491. void CMasterGraph::done()
  2492. {
  2493. if (started)
  2494. {
  2495. if (!aborted && (!queryOwner() || isGlobal()))
  2496. getFinalProgress(); // waiting for slave graph to finish and send final progress update + extra act end info.
  2497. }
  2498. CGraphBase::done();
  2499. if (started && !queryOwner())
  2500. {
  2501. if (globals->getPropBool("@watchdogProgressEnabled"))
  2502. queryJobManager().queryDeMonServer()->endGraph(this, true);
  2503. }
  2504. }
  2505. void CMasterGraph::setComplete(bool tf)
  2506. {
  2507. CGraphBase::setComplete(tf);
  2508. if (tf && !queryOwner())
  2509. {
  2510. job.queryWorkUnit().setNodeState(job.queryGraphName(), graphId, graphDone?WUGraphComplete:WUGraphFailed);
  2511. }
  2512. }
  2513. bool CMasterGraph::deserializeStats(unsigned node, MemoryBuffer &mb)
  2514. {
  2515. CriticalBlock b(createdCrit);
  2516. unsigned count, _count;
  2517. mb.read(count);
  2518. if (count)
  2519. setProgressUpdated();
  2520. _count = count;
  2521. while (count--)
  2522. {
  2523. activity_id activityId;
  2524. mb.read(activityId);
  2525. CMasterActivity *activity = NULL;
  2526. CMasterGraphElement *element = (CMasterGraphElement *)queryElement(activityId);
  2527. if (element)
  2528. {
  2529. activity = (CMasterActivity *)element->queryActivity();
  2530. if (!activity)
  2531. {
  2532. CGraphBase *parentGraph = element->queryOwner().queryOwner(); // i.e. am I in a childgraph
  2533. if (!parentGraph)
  2534. {
  2535. GraphPrintLog("Activity id=%" ACTPF "d not created in master and not a child query activity", activityId);
  2536. return false; // don't know if or how this could happen, but all bets off with packet if did.
  2537. }
  2538. Owned<IException> e;
  2539. try
  2540. {
  2541. element->createActivity();
  2542. activity = (CMasterActivity *)element->queryActivity();
  2543. }
  2544. catch (IException *_e)
  2545. {
  2546. e.setown(_e);
  2547. GraphPrintLog(_e, "In deserializeStats");
  2548. }
  2549. if (!activity || e.get())
  2550. {
  2551. GraphPrintLog("Activity id=%" ACTPF "d failed to created child query activity ready for progress", activityId);
  2552. return false;
  2553. }
  2554. }
  2555. if (activity)
  2556. activity->deserializeStats(node, mb);
  2557. }
  2558. else
  2559. {
  2560. GraphPrintLog("Failed to find activity, during progress deserialization, id=%" ACTPF "d", activityId);
  2561. return false; // don't know if or how this could happen, but all bets off with packet if did.
  2562. }
  2563. }
  2564. unsigned subs, _subs;
  2565. mb.read(subs);
  2566. _subs = subs;
  2567. while (subs--)
  2568. {
  2569. graph_id subId;
  2570. mb.read(subId);
  2571. Owned<CMasterGraph> graph = (CMasterGraph *)job.queryJobChannel(0).getGraph(subId);
  2572. if (NULL == graph.get())
  2573. return false;
  2574. if (!graph->deserializeStats(node, mb))
  2575. return false;
  2576. }
  2577. return true;
  2578. }
  2579. IThorResult *CMasterGraph::createResult(CActivityBase &activity, unsigned id, IThorGraphResults *results, IThorRowInterfaces *rowIf, ThorGraphResultType resultType, unsigned spillPriority)
  2580. {
  2581. Owned<CCollatedResult> result = new CCollatedResult(*this, activity, rowIf, id, results->queryOwnerId(), spillPriority);
  2582. results->setResult(id, result);
  2583. return result;
  2584. }
  2585. IThorResult *CMasterGraph::createResult(CActivityBase &activity, unsigned id, IThorRowInterfaces *rowIf, ThorGraphResultType resultType, unsigned spillPriority)
  2586. {
  2587. Owned<CCollatedResult> result = new CCollatedResult(*this, activity, rowIf, id, localResults->queryOwnerId(), spillPriority);
  2588. localResults->setResult(id, result);
  2589. return result;
  2590. }
  2591. IThorResult *CMasterGraph::createGraphLoopResult(CActivityBase &activity, IThorRowInterfaces *rowIf, ThorGraphResultType resultType, unsigned spillPriority)
  2592. {
  2593. Owned<CCollatedResult> result = new CCollatedResult(*this, activity, rowIf, 0, localResults->queryOwnerId(), spillPriority);
  2594. unsigned id = graphLoopResults->addResult(result);
  2595. result->setId(id);
  2596. return result;
  2597. }
  2598. ///////////////////////////////////////////////////
  2599. static bool suppressStatisticIfZero(StatisticKind kind)
  2600. {
  2601. switch (kind)
  2602. {
  2603. case StNumSpills:
  2604. case StSizeSpillFile:
  2605. case StTimeSpillElapsed:
  2606. case StNumDiskRetries:
  2607. return true;
  2608. }
  2609. return false;
  2610. }
  2611. ///////////////////////////////////////////////////
  2612. CThorStats::CThorStats(CJobBase &_ctx, StatisticKind _kind) : ctx(_ctx), kind(_kind)
  2613. {
  2614. unsigned c = queryClusterWidth();
  2615. while (c--) counts.append(0);
  2616. reset();
  2617. }
  2618. void CThorStats::extract(unsigned node, const CRuntimeStatisticCollection & stats)
  2619. {
  2620. set(node, stats.getStatisticValue(kind));
  2621. }
  2622. void CThorStats::set(unsigned node, unsigned __int64 count)
  2623. {
  2624. counts.replace(count, node);
  2625. }
  2626. void CThorStats::reset()
  2627. {
  2628. tot = max = avg = 0;
  2629. min = (unsigned __int64) -1;
  2630. minNode = maxNode = maxSkew = minSkew = maxNode = minNode = 0;
  2631. }
  2632. void CThorStats::calculateSkew()
  2633. {
  2634. unsigned count = counts.ordinality();
  2635. double _avg = (double)tot/count;
  2636. if (_avg)
  2637. {
  2638. if (max > ctx.querySlaves()) // i.e. if small count, suppress skew stats.
  2639. {
  2640. //MORE: Range protection on maxSkew?
  2641. maxSkew = (unsigned)(10000.0 * (((double)max-_avg)/_avg));
  2642. minSkew = (unsigned)(10000.0 * ((_avg-(double)min)/_avg));
  2643. }
  2644. avg = (unsigned __int64)_avg;
  2645. }
  2646. }
  2647. void CThorStats::tallyValue(unsigned __int64 thiscount, unsigned n)
  2648. {
  2649. tot += thiscount;
  2650. if (thiscount > max)
  2651. {
  2652. max = thiscount;
  2653. maxNode = n;
  2654. }
  2655. if (thiscount < min)
  2656. {
  2657. min = thiscount;
  2658. minNode = n;
  2659. }
  2660. }
  2661. void CThorStats::processInfo()
  2662. {
  2663. reset();
  2664. ForEachItemIn(n, counts)
  2665. {
  2666. unsigned __int64 thiscount = counts.item(n);
  2667. tallyValue(thiscount, n+1);
  2668. }
  2669. calculateSkew();
  2670. }
  2671. void CThorStats::getStats(IStatisticGatherer & stats, bool suppressMinMaxWhenEqual)
  2672. {
  2673. processInfo();
  2674. if ((0 == tot) && suppressStatisticIfZero(kind))
  2675. return;
  2676. //MORE: For most measures (not time stamps etc.) it would be sensible to output the total here....
  2677. if (!suppressMinMaxWhenEqual || (maxSkew != minSkew))
  2678. {
  2679. stats.addStatistic((StatisticKind)(kind|StMinX), min);
  2680. stats.addStatistic((StatisticKind)(kind|StMaxX), max);
  2681. stats.addStatistic((StatisticKind)(kind|StAvgX), avg);
  2682. }
  2683. if (maxSkew != minSkew)
  2684. {
  2685. stats.addStatistic((StatisticKind)(kind|StSkewMin), -(__int64)minSkew); // Save minimum as a negative value so consistent
  2686. stats.addStatistic((StatisticKind)(kind|StSkewMax), maxSkew);
  2687. stats.addStatistic((StatisticKind)(kind|StNodeMin), minNode);
  2688. stats.addStatistic((StatisticKind)(kind|StNodeMax), maxNode);
  2689. }
  2690. }
  2691. ///////////////////////////////////////////////////
  2692. CTimingInfo::CTimingInfo(CJobBase &ctx) : CThorStats(ctx, StTimeLocalExecute)
  2693. {
  2694. }
  2695. ///////////////////////////////////////////////////
  2696. ProgressInfo::ProgressInfo(CJobBase &ctx) : CThorStats(ctx, StNumRowsProcessed)
  2697. {
  2698. startcount = stopcount = 0;
  2699. }
  2700. void ProgressInfo::processInfo() // reimplement as counts have special flags (i.e. stop/start)
  2701. {
  2702. reset();
  2703. startcount = stopcount = 0;
  2704. ForEachItemIn(n, counts)
  2705. {
  2706. unsigned __int64 thiscount = counts.item(n);
  2707. if (thiscount & THORDATALINK_STARTED)
  2708. startcount++;
  2709. if (thiscount & THORDATALINK_STOPPED)
  2710. stopcount++;
  2711. thiscount = thiscount & THORDATALINK_COUNT_MASK;
  2712. tallyValue(thiscount, n+1);
  2713. }
  2714. calculateSkew();
  2715. }
  2716. void ProgressInfo::getStats(IStatisticGatherer & stats)
  2717. {
  2718. CThorStats::getStats(stats, true);
  2719. stats.addStatistic(kind, tot);
  2720. stats.addStatistic(StNumSlaves, counts.ordinality());
  2721. stats.addStatistic(StNumStarts, startcount);
  2722. stats.addStatistic(StNumStops, stopcount);
  2723. }
  2724. ///////////////////////////////////////////////////
  2725. CJobMaster *createThorGraph(const char *graphName, IConstWorkUnit &workunit, ILoadedDllEntry *querySo, bool sendSo, const SocketEndpoint &agentEp)
  2726. {
  2727. Owned<CJobMaster> jobMaster = new CJobMaster(workunit, graphName, querySo, sendSo, agentEp);
  2728. IPropertyTree *graphXGMML = jobMaster->queryGraphXGMML();
  2729. Owned<IPropertyTreeIterator> iter = graphXGMML->getElements("node");
  2730. ForEach(*iter)
  2731. jobMaster->addSubGraph(iter->query());
  2732. jobMaster->addDependencies(graphXGMML);
  2733. return LINK(jobMaster);
  2734. }
  2735. static IJobManager *jobManager = NULL;
  2736. void setJobManager(IJobManager *_jobManager)
  2737. {
  2738. CriticalBlock b(*jobManagerCrit);
  2739. jobManager = _jobManager;
  2740. }
  2741. IJobManager *getJobManager()
  2742. {
  2743. CriticalBlock b(*jobManagerCrit);
  2744. return LINK(jobManager);
  2745. }
  2746. IJobManager &queryJobManager()
  2747. {
  2748. CriticalBlock b(*jobManagerCrit);
  2749. assertex(jobManager);
  2750. return *jobManager;
  2751. }