XRootD
XrdPfcResourceMonitor.cc
Go to the documentation of this file.
2 #include "XrdPfc.hh"
4 #include "XrdPfcFsTraversal.hh"
5 #include "XrdPfcDirState.hh"
8 #include "XrdPfcTrace.hh"
9 #include "XrdPfcPurgePin.hh"
10 
11 #include "XrdOss/XrdOss.hh"
12 
13 #include <algorithm>
14 
15 // #define RM_DEBUG
16 #ifdef RM_DEBUG
17 #define dprintf(...) printf(__VA_ARGS__)
18 #else
19 #define dprintf(...) (void(0))
20 #endif
21 
22 using namespace XrdPfc;
23 
24 namespace
25 {
26  XrdSysTrace* GetTrace() { return Cache::GetInstance().GetTrace(); }
27  const char *m_traceID = "ResourceMonitor";
28 }
29 
30 //------------------------------------------------------------------------------
31 
33  m_fs_state(* new DataFsState),
34  m_oss(oss)
35 {}
36 
38 {
39  delete &m_fs_state;
40 }
41 
42 //------------------------------------------------------------------------------
43 // Initial scan
44 //------------------------------------------------------------------------------
45 
47 {
48  m_dir_scan_mutex.Lock();
49  if (m_dir_scan_in_progress) {
50  bool dir_checked = false;
51  m_dir_scan_open_requests.push_back({lfn, cond, dir_checked});
52  cond.Lock();
53  m_dir_scan_mutex.UnLock();
54  while ( ! dir_checked)
55  cond.Wait();
56  cond.UnLock();
57  } else {
58  m_dir_scan_mutex.UnLock();
59  }
60 }
61 
62 void ResourceMonitor::check_inter_dir_scan_open_requests()
63 {
64  // Called from the scan thread between directories and periodically while reading
65  // a large directory. Rate-limit on time rather than on number of directories
66  // visited: a single directory can take minutes to stat on a large cache.
67  auto now = std::chrono::steady_clock::now();
68  if (now - m_dir_scan_last_check < s_dir_scan_check_interval)
69  return;
70  m_dir_scan_last_check = now;
71 
72  process_inter_dir_scan_open_requests();
73 }
74 
75 void ResourceMonitor::process_inter_dir_scan_open_requests()
76 {
77  m_dir_scan_mutex.Lock();
78  // Only service the requests present on entry. Under a steady stream of opens the
79  // queue may never become empty and the initial scan would be starved.
80  size_t n_to_process = m_dir_scan_open_requests.size();
81  if (n_to_process == 0)
82  {
83  m_dir_scan_mutex.UnLock();
84  return;
85  }
86  // Use a private traversal object: this may be called while the scan's traversal
87  // is in the middle of reading a directory.
88  FsTraversal fst(m_oss);
89  while (n_to_process-- > 0)
90  {
91  LfnCondRecord &lcr = m_dir_scan_open_requests.front();
92  m_dir_scan_mutex.UnLock();
93 
94  cross_check_or_process_oob_lfn(lcr.f_lfn, fst);
95  lcr.f_cond.Lock();
96  lcr.f_checked = true;
97  lcr.f_cond.Signal();
98  lcr.f_cond.UnLock();
99 
100  m_dir_scan_mutex.Lock();
101  m_dir_scan_open_requests.pop_front();
102  }
103  m_dir_scan_mutex.UnLock();
104 }
105 
106 void ResourceMonitor::cross_check_or_process_oob_lfn(const std::string &lfn, FsTraversal &fst)
107 {
108  // Check if lfn has already been processed ... or process it now and mark
109  // the DirState accordingly (partially processed oob).
110  static const char *trc_pfx = "cross_check_or_process_oob_lfn() ";
111 
112  // last-existing-dir-state is currently not used. It was expected some work would need
113  // to be done along the path ... but we know they are all empty as they were just
114  // created. We could tag them all as already scanned.
115  DirState *last_existing_ds = nullptr;
116  DirState *ds = m_fs_state.find_dirstate_for_lfn(lfn, &last_existing_ds);
117  if (ds->m_scanned)
118  return;
119 
120  size_t pos = lfn.find_last_of("/");
121  std::string dir = (pos == std::string::npos) ? "" : lfn.substr(0, pos);
122 
123  XrdOssDF *dhp = m_oss.newDir(trc_pfx);
124  bool opened = (dhp->Opendir(dir.c_str(), fst.default_env()) == XrdOssOK);
125  if (opened)
126  {
127  fst.slurp_dir_ll(*dhp, ds->m_depth, dir.c_str(), trc_pfx);
128 
129  // XXXX clone of function below .... move somewhere? Esp. removal of non-paired files?
130  DirUsage &here = ds->m_here_usage;
131  for (auto it = fst.m_current_files.begin(); it != fst.m_current_files.end(); ++it)
132  {
133  if (it->second.has_data && it->second.has_cinfo) {
134  here.m_StBlocks += it->second.stat_data.st_blocks;
135  here.m_NFiles += 1;
136  }
137  }
138  }
139  delete dhp;
140  // Only a directory that was read counts as scanned, else its files are skipped.
141  if (opened)
142  ds->m_scanned = true;
143 }
144 
146 {
147  dprintf("In scan_dir_and_recurse for '%s', size of dir_vec = %d, file_stat_map = %d\n",
148  fst.m_current_path.c_str(),
149  (int)fst.m_current_dirs.size(), (int)fst.m_current_files.size());
150 
151  // Breadth first, accumulate into "here", unless it was already scanned via an
152  // OOB open file request.
153  if ( ! fst.m_dir_state->m_scanned)
154  {
155  DirUsage &here = fst.m_dir_state->m_here_usage;
156  for (auto it = fst.m_current_files.begin(); it != fst.m_current_files.end(); ++it)
157  {
158  dprintf("would be doing something with %s ... has_data=%d, has_cinfo=%d\n",
159  it->first.c_str(), it->second.has_data, it->second.has_cinfo);
160 
161  // XXX Make some of these optional?
162  // Remove files that do not have both cinfo and data?
163  // Remove empty directories before even descending?
164  // Leave this for some consistency pass?
165  // Note that FsTraversal supports ignored paths ... some details (config, N2N) to be clarified.
166 
167  if (it->second.has_data && it->second.has_cinfo) {
168  here.m_StBlocks += it->second.stat_data.st_blocks;
169  here.m_NFiles += 1;
170  }
171  }
172  fst.m_dir_state->m_scanned = true;
173  }
174 
175  // Swap-out directories as recursion reuses the FsTraversal.
176  std::vector<std::string> dirs;
177  dirs.swap(fst.m_current_dirs);
178 
179  check_inter_dir_scan_open_requests();
180 
181  // Descend into sub-dirs, do not accumulate into recursive_subdir_usage yet. This is done
182  // in a separate pass to allow for proper accounting of files being opened during the initial scan.
183  for (auto &dname : dirs)
184  {
185  if (fst.cd_down(dname))
186  {
188  fst.cd_up();
189  }
190  // XXX else try to remove it?
191  }
192 }
193 
195 {
196  // Called after PFC configuration is complete, but before full startup of the daemon.
197  // Base line usages are accumulated as part of the file-system, traversal.
198 
200 
201  DirState *root_ds = m_fs_state.get_root();
202  FsTraversal fst(m_oss);
203  fst.m_protected_top_dirs.insert("pfc-stats"); // XXXX This should come from config. Also: N2N?
204  // Keep servicing file-open requests while reading huge directories.
205  fst.m_readdir_hook = [this]() { check_inter_dir_scan_open_requests(); };
206 
207  if ( ! fst.begin_traversal(root_ds, "/"))
208  return false;
209 
210  // m_dir_scan_in_progress is initialized to true in ResourceMonitor.hh to avoid a race at startup.
211 
213 
214  fst.end_traversal();
215 
216  // We have all directories scanned, available in DirState tree, let all remaining files go
217  // and then we shall do the upward propagation of usages.
218  {
219  XrdSysMutexHelper _lock(m_dir_scan_mutex);
220  m_dir_scan_in_progress = false;
221  }
222  // m_dir_scan_open_requests should now be final, ie, no new entries will be added.
223  while ( ! m_dir_scan_open_requests.empty())
224  {
225  LfnCondRecord &lcr = m_dir_scan_open_requests.front();
226  lcr.f_cond.Lock();
227  lcr.f_checked = true;
228  lcr.f_cond.Signal();
229  lcr.f_cond.UnLock();
230 
231  m_dir_scan_open_requests.pop_front();
232  }
233 
234  // Do upward propagation of usages.
236  m_current_usage_in_st_blocks = root_ds->m_here_usage.m_StBlocks +
239 
240  return true;
241 }
242 
243 //------------------------------------------------------------------------------
244 // Processing of queues
245 //------------------------------------------------------------------------------
246 
248 {
249  static const char *trc_pfx = "process_queues() ";
250 
251  // Assure that we pick up only entries that are present now.
252  // We really want all open records to be processed before file-stats updates
253  // and all those before the close records.
254  // Purges are sort of tangential as they really just modify bytes / number
255  // of files in a directory and do not deal with any persistent file id tokens.
256 
257  int n_records = 0;
258  {
259  XrdSysMutexHelper _lock(&m_queue_mutex);
260  n_records += m_file_open_q.swap_queues();
261  n_records += m_file_update_stats_q.swap_queues();
262  n_records += m_file_close_q.swap_queues();
263  n_records += m_file_purge_q1.swap_queues();
264  n_records += m_file_purge_q2.swap_queues();
265  n_records += m_file_purge_q3.swap_queues();
266  ++m_queue_swap_u1;
267  }
268 
269  for (auto &i : m_file_open_q.read_queue())
270  {
271  // i.id: LFN, i.record: OpenRecord
272  AccessToken &at = token(i.id);
273  dprintf("process file open for token %d, time %ld -- %s\n",
274  i.id, i.record.m_open_time, at.m_filename.c_str());
275 
276  // Resolve fname into DirState.
277  // We could clear the filename after this ... or keep it, should we need it later on.
278  // For now it is just used for debug printouts.
279  DirState *last_existing_ds = nullptr;
280  DirState *ds = m_fs_state.find_dirstate_for_lfn(at.m_filename, &last_existing_ds);
281  at.m_dir_state = ds;
282  note_file_open(ds, last_existing_ds, i.record.m_existing_file);
283 
284  ds->m_here_usage.m_LastOpenTime = i.record.m_open_time;
285  }
286 
287  for (auto &i : m_file_update_stats_q.read_queue())
288  {
289  // i.id: token, i.record: Stats
290  AccessToken &at = token(i.id);
291  // Stats
292  DirState *ds = at.m_dir_state;
293  dprintf("process file update for token %d, %p -- %s\n",
294  i.id, ds, at.m_filename.c_str());
295 
296  ds->m_here_stats.AddUp(i.record);
297  m_current_usage_in_st_blocks += i.record.m_StBlocksAdded;
298  }
299 
300  for (auto &i : m_file_close_q.read_queue())
301  {
302  // i.id: token, i.record: CloseRecord
303  AccessToken &at = token(i.id);
304  dprintf("process file close for token %d, time %ld -- %s\n",
305  i.id, i.record.m_close_time, at.m_filename.c_str());
306 
307  DirState *ds = at.m_dir_state;
308  ds->m_here_stats.m_NFilesClosed += 1;
309  ds->m_here_usage.m_LastCloseTime = i.record.m_close_time;
310 
311  }
312  { // Under the lock: clear() writes fields register_file_update_stats() uses.
313  XrdSysMutexHelper _lock(&m_queue_mutex);
314  for (auto &i : m_file_close_q.read_queue())
315  m_access_tokens.release(i.id);
316  }
317 
318  for (auto &i : m_file_purge_q1.read_queue())
319  {
320  // i.id: DirState*, i.record: PurgeRecord
321  DirState *ds = i.id;
322  ds->m_here_stats.m_StBlocksRemoved += i.record.m_size_in_st_blocks;
323  ds->m_here_stats.m_NFilesRemoved += i.record.m_n_files;
324  m_current_usage_in_st_blocks -= i.record.m_size_in_st_blocks;
325  }
326  for (auto &i : m_file_purge_q2.read_queue())
327  {
328  // i.id: directory-path, i.record: PurgeRecord
329  DirState *ds = m_fs_state.get_root()->find_path(i.id, -1, false, false);
330  if ( ! ds) {
331  TRACE(Error, trc_pfx << "DirState not found for directory path '" << i.id << "'.");
332  // find_path can return the last dir found ... but this clearly isn't a valid purge record.
333  continue;
334  }
335  ds->m_here_stats.m_StBlocksRemoved += i.record.m_size_in_st_blocks;
336  ds->m_here_stats.m_NFilesRemoved += i.record.m_n_files;
337  m_current_usage_in_st_blocks -= i.record.m_size_in_st_blocks;
338  }
339  for (auto &i : m_file_purge_q3.read_queue())
340  {
341  // i.id: LFN, i.record: size of file in st_blocks
342  DirState *ds = m_fs_state.get_root()->find_path(i.id, -1, true, false);
343  if ( ! ds) {
344  TRACE(Error, trc_pfx << "DirState not found for LFN path '" << i.id << "'.");
345  continue;
346  }
347  ds->m_here_stats.m_StBlocksRemoved += i.record;
348  ds->m_here_stats.m_NFilesRemoved += 1;
349  m_current_usage_in_st_blocks -= i.record;
350  }
351 
352  // Read queues / vectors are cleared at swap time.
353  // We might consider reducing their capacity by half if, say, their usage is below 25%.
354 
355  return n_records;
356 }
357 
358 //------------------------------------------------------------------------------
359 // Heart beat
360 //------------------------------------------------------------------------------
361 
363 {
364  static const char *tpfx = "heart_beat() ";
365 
366  const Configuration &conf = Cache::Conf();
367 
368  const int s_queue_proc_interval = 10;
369  const int s_sshot_report_interval = conf.m_dirStatsInterval; // one of 60, 120, 300,
370  // 600, 900, 1200, 1800, 3600 s
371  const int s_purge_check_interval = 60;
372  const int s_purge_report_interval = conf.m_purgeInterval;
373  const int s_purge_cold_files_interval = conf.m_purgeInterval * conf.m_purgeAgeBasedPeriod;
374 
375  // initial scan performed as part of config
376 
377  time_t now = time(0);
378  time_t next_queue_proc_time = now + s_queue_proc_interval;
379  time_t next_sshot_report_time = (now / 60) * 60 + 60; // at next full minute
380  time_t next_purge_check_time = now + s_purge_check_interval;
381  time_t next_purge_report_time = now + s_purge_report_interval;
382  time_t next_purge_cold_files_time = now + s_purge_cold_files_interval;
383 
384  long long n_reported_token_failures = 0;
385 
386  while (true)
387  {
388  time_t start = time(0);
389  time_t next_event = std::min({ next_queue_proc_time, next_sshot_report_time,
390  next_purge_check_time, next_purge_report_time, next_purge_cold_files_time });
391 
392  if (next_event > start)
393  {
394  unsigned int t_sleep = next_event - start;
395  TRACE(Dump, tpfx << "sleeping for " << t_sleep << " seconds until the next beat.");
396  sleep(t_sleep);
397  }
398 
399  // Check if purge has been running and has completed yet.
400  // For now this is only used to prevent removal of empty leaf directories
401  // during stat propagation so we do not need to wait for the condition in
402  // the above sleep.
403  if (m_purge_task_active) {
405  if (m_purge_task_complete) {
407  }
408  }
409 
410  time_t queue_swap_time = time(0);
411 
412  // Always process the queues.
413  int n_processed = process_queues();
414  next_queue_proc_time = queue_swap_time + s_queue_proc_interval;
415  TRACE(Dump, tpfx << "process_queues -- n_records=" << n_processed);
416 
417  { // Access-token space exhausted -- should be unreachable, so say so loudly.
418  long long taf;
419  { XrdSysMutexHelper _lock(&m_queue_mutex); taf = m_token_alloc_failures; }
420  if (taf > n_reported_token_failures) {
421  TRACE(Error, tpfx << "access-token space exhausted "
422  << taf - n_reported_token_failures << " time(s) (total " << taf
423  << "); those files are open but not accounted for. Raise MaxChunks "
424  "on ResourceMonitor::m_access_tokens.");
425  n_reported_token_failures = taf;
426  }
427  }
428 
429  // Always update basic info on m_fs_state (space, usage, file_usage).
431 
432  now = time(0);
433 
434  // Make planning for fs_state_update, sshot dump and purge task.
435  // Second two require the first, so figure out what is going to happen.
436  bool do_sshot_report = next_sshot_report_time <= now;
437  bool do_purge_check = next_purge_check_time <= now;
438  bool do_purge_report = next_purge_report_time <= now;
439  bool do_purge_cold_files = next_purge_cold_files_time <= now;
440 
441  // Update stats in usages if any secondary activity will happen.
442  if (do_sshot_report || do_purge_check || do_purge_report || do_purge_cold_files)
443  {
444  unlink_func unlink_foo = [&](const std::string &dp)->int {
445  int ret = m_oss.Unlink(dp.c_str());
446  if (ret != 0) {
447  TRACE(Info, tpfx << "Empty dir unlink error: " << ret << " at " << dp);
448  } else {
449  TRACE(Debug, tpfx << "Empty dir unlink success: " << dp);
450  }
451  return ret;
452  };
453 
454  // Potentially prune the empty leaf dirs even less frequently, once per hour, maybe?
455  bool purge_leaf_dirs = do_sshot_report && ! m_purge_task_active;
456  m_fs_state.update_stats_and_usages(queue_swap_time, purge_leaf_dirs, unlink_foo);
457 
458  // This reporting into log/stdout is to be removed.
459  // Meaning of conf.is_dir_stat_reporting_on() etc is to be clarified / improved.
460  if (do_sshot_report && conf.is_dir_stat_reporting_on())
461  {
462  const int store_depth = conf.m_dirStatsStoreDepth;
463  #ifdef RM_DEBUG
464  const DirState &root_ds = *m_fs_state.get_root();
465  dprintf("Snapshot n_dirs=%d, total n_dirs=%d\n", root_ds.count_dirs_to_level(store_depth),
467  #endif
468  m_fs_state.dump_recursively(store_depth);
469  }
470 
471  m_fs_state.reset_stats(queue_swap_time);
472  }
473 
474  if (do_sshot_report)
475  {
476  // Sshot reports are equidistant, at "full" reporting interval.
477  next_sshot_report_time = ((now + 1) / s_sshot_report_interval) * s_sshot_report_interval + s_sshot_report_interval;
478 
479  // This should dump out binary snapshot into /pfc-stats/, if so configured.
480 
481  // json dump to std::out for debug purpose
482  DataFsSnapshot ss(m_fs_state, m_fs_state.m_sshot_stats_reset_time);
483  const DirState &root_ds = *m_fs_state.get_root();
484  const int store_depth = conf.m_dirStatsStoreDepth;
485  const int n_sshot_dirs = root_ds.count_dirs_to_level(store_depth);
486  ss.m_dir_states.reserve(n_sshot_dirs);
487  ss.m_dir_states.emplace_back( DirStateElement(root_ds, -1) );
488  fill_sshot_vec_children(root_ds, 0, ss.m_dir_states, store_depth);
489 
490  // This should really be export to a file (preferably binary, but then bin->json command is needed, too).
491  // ss.dump();
492 
493  const char* dumpfile = "/pfc-stats/DirStat.json";
494  ss.write_json_file(dumpfile, m_oss, false);
495  m_fs_state.reset_sshot_stats(queue_swap_time);
496  }
497 
498  if (do_purge_check || do_purge_report || do_purge_cold_files)
499  {
500  perform_purge_check(do_purge_cold_files, do_purge_report ? TRACE_Info : TRACE_Debug);
501 
502  next_purge_check_time = now + s_purge_check_interval;
503  if (do_purge_report) next_purge_report_time = now + s_purge_report_interval;
504  if (do_purge_cold_files) next_purge_cold_files_time = now + s_purge_cold_files_interval;
505  }
506 
507  } // end while forever
508 }
509 
510 //------------------------------------------------------------------------------
511 // DirState export helpers
512 //------------------------------------------------------------------------------
513 
515  int parent_idx,
516  std::vector<DirStateElement> &vec,
517  int max_depth)
518 {
519  const int pos0 = (int) vec.size();
520  const int n_children = (int) parent_ds.m_subdirs.size();
521 
522  // maxdepth 0 is the root alone: a node at max_depth is emitted, its children are not.
523  if (parent_ds.m_depth >= max_depth)
524  {
525  // Truncated here: keep the daughter range at (-1, -1). No children: empty range.
526  if (n_children == 0) {
527  DirStateElement &dse = vec[parent_idx];
528  dse.m_daughters_begin = dse.m_daughters_end = pos0;
529  }
530  return;
531  }
532 
533  { // scoped: the emplace_back loop below can reallocate vec
534  DirStateElement &parent_dse = vec[parent_idx];
535  parent_dse.m_daughters_begin = pos0;
536  parent_dse.m_daughters_end = pos0 + n_children;
537  }
538 
539  if (n_children == 0) return;
540 
541  for (auto const & [name, child] : parent_ds.m_subdirs)
542  {
543  vec.emplace_back( DirStateElement(child, parent_idx) );
544  }
545 
546  int pos = pos0;
547  for (auto const & [name, child] : parent_ds.m_subdirs)
548  {
549  fill_sshot_vec_children(child, pos, vec, max_depth);
550  ++pos;
551  }
552 }
553 
555  int parent_idx,
556  std::vector<DirPurgeElement> &vec,
557  int max_depth)
558 {
559  const int pos0 = (int) vec.size();
560  const int n_children = (int) parent_ds.m_subdirs.size();
561 
562  // maxdepth 0 is the root alone: a node at max_depth is emitted, its children are not.
563  if (parent_ds.m_depth >= max_depth)
564  {
565  // Truncated here: keep the daughter range at (-1, -1). No children: empty range.
566  if (n_children == 0) {
567  DirPurgeElement &dse = vec[parent_idx];
568  dse.m_daughters_begin = dse.m_daughters_end = pos0;
569  }
570  return;
571  }
572 
573  { // scoped: the emplace_back loop below can reallocate vec
574  DirPurgeElement &parent_dse = vec[parent_idx];
575  parent_dse.m_daughters_begin = pos0;
576  parent_dse.m_daughters_end = pos0 + n_children;
577  }
578 
579  if (n_children == 0) return;
580 
581  for (auto const & [name, child] : parent_ds.m_subdirs)
582  {
583  vec.emplace_back( DirPurgeElement(child, child.m_here_usage, child.m_recursive_subdir_usage, parent_idx) );
584  }
585 
586  int pos = pos0;
587  for (auto const & [name, child] : parent_ds.m_subdirs)
588  {
589  fill_pshot_vec_children(child, pos, vec, max_depth);
590  ++pos;
591  }
592 }
593 
594 //------------------------------------------------------------------------------
595 // Purge helpers, drivers, etc.
596 //------------------------------------------------------------------------------
597 
599 {
600  static const char *trc_pfx = "update_vs_and_file_usage_info() ";
601 
602  const auto &conf = Cache::Conf();
603  XrdOssVSInfo vsi;
604 
605  // StatVS error (after it succeeded in config) implies a memory corruption (according to Mr. H).
606  if (m_oss.StatVS(&vsi, conf.m_data_space.c_str(), 1) < 0) {
607  TRACE(Error, trc_pfx << "can't get StatVS for oss space '" << conf.m_data_space << "'. This is a fatal error.");
608  _exit(1);
609  }
610  m_fs_state.m_disk_total = vsi.Total;
611  m_fs_state.m_disk_used = vsi.Total - vsi.Free;
612  m_fs_state.m_file_usage = 512ll * m_current_usage_in_st_blocks;
613  if (m_oss.StatVS(&vsi, conf.m_meta_space.c_str(), 1) < 0) {
614  TRACE(Error, trc_pfx << "can't get StatVS for oss space '" << conf.m_meta_space << "'. This is a fatal error.");
615  _exit(1);
616  }
617  m_fs_state.m_meta_total = vsi.Total;
618  m_fs_state.m_meta_used = vsi.Total - vsi.Free;
619 }
620 
621 long long ResourceMonitor::get_file_usage_bytes_to_remove(const DataFsPurgeshot &ps, long long write_estimate, int tl)
622 {
623  // short names from config values
624  const Configuration &conf = Cache::Conf();
625  long long f0 = conf.m_fileUsageBaseline;
626  long long f1 = conf.m_fileUsageNominal;
627  long long f2 = conf.m_fileUsageMax;
628  long long w1 = conf.m_diskUsageLWM;
629  long long w2 = conf.m_diskUsageHWM;
630 
631  // get usage from purge snapshot
632  long long T = ps.m_disk_total;
633  long long x = ps.m_file_usage;
634  long long u = ps.m_disk_used;
635 
636  // get file usage increase from the previous time interval check
637  long long delta = write_estimate;
638  TRACE_INT(tl, "file usage increased since the previous purge interval in bytes: " << delta );
639 
640  long long bytes_to_remove = 0;
641 
642  // helper lambda function
643  auto clamp = [&x, &bytes_to_remove](long long lowval, long long highval)
644  {
645  long long val = x;
646  long long newval = val - bytes_to_remove;
647 
648  // removed too much
649  if (newval < lowval)
650  {
651  return lowval - val;
652  }
653 
654  // removed too little
655  if (newval > highval)
656  {
657  return val - highval;
658  }
659  // keep the original value
660  return bytes_to_remove;
661  };
662 
663  // under file quota, nothing to do
664  if (x < f0)
665  return 0;
666 
667  // total disk usage exceeds highWatermark
668  if (u >= w2)
669  {
670  TRACE_INT(tl, "Disk usage: " << ps.m_disk_used << " exceed highWatermark " << conf.m_diskUsageHWM);
671  float frac_u = static_cast<float>(u - w2) / (T - w2);
672  float frac_x = static_cast<float>(x - f0) / (f1 - f0);
673 
674  if (w2 == T)
675  {
676  bytes_to_remove = u -w1;
677  }
678  else
679  {
680  if (frac_x > frac_u)
681  {
682  // the cache is the reason for going out of w2 range
683  bytes_to_remove = (frac_x - frac_u) * (f1 - f0);
684  bytes_to_remove += delta;
685  bytes_to_remove = clamp(f0, f1);
686  }
687  else
688  {
689  // someone else is filling disk space, go to f1
690  bytes_to_remove = clamp(f0, f2);
691  }
692  return bytes_to_remove;
693  }
694  }
695 
696  // file quota and total disk usage is within normal range, check if this space usage is
697  // proportinal to disk usage and correct it
698  if (u > w1 && x > f1)
699  {
700  float frac_u = static_cast<float>(u - w1) / (w2 - w1);
701  float frac_x = static_cast<float>(x - f1) / (f2 - f1);
702  if (frac_x > frac_u)
703  {
704  TRACE_INT(tl, "Disproportional file quota usage comapared to disc usage (frac_x/frac_u) = " << frac_x << "/"<< frac_u);
705  bytes_to_remove = (frac_x - frac_u) * (f2 - f1);
706  bytes_to_remove += delta;
707  }
708 
709  // check the new x val will not be below f0
710  bytes_to_remove = clamp(f0, f2);
711  return bytes_to_remove;
712  }
713 
714  // final check: disk useage is lower that w1, check if exceed the max file usage f2
715  if (x > f2)
716  {
717  // drop usage to f2
718  // compare with global disk usage in the previous purge cycle (default 300s)
719  // check delta is not overflowing f2, else set numver of bytes to remove according remove to f0
720 
721  TRACE_INT(tl, "File usage exceeds maxim file usage. Total disk usage is under lowWatermark. Clearing to low file usage.");
722  long long f2delta = std::max(f2 - delta, f0);
723  bytes_to_remove = clamp(f0, f2delta);
724  return bytes_to_remove;
725  }
726 
727  return bytes_to_remove;
728 }
729 
730 void ResourceMonitor::perform_purge_check(bool purge_cold_files, int tl)
731 {
732  static const char *trc_pfx = "perform_purge_check() ";
733  const Configuration &conf = Cache::Conf();
734 
735  std::unique_ptr<DataFsPurgeshot> psp( new DataFsPurgeshot(m_fs_state) );
736  DataFsPurgeshot &ps = *psp;
737 
738  ps.m_file_usage = 512ll * m_current_usage_in_st_blocks;
739  // These are potentially wrong as cache might be writing over preallocated byte ranges.
741  // Can have another estimate based on eiter writes or st-blocks from purge-stats, once we have them.
742 
743  TRACE_INT(tl, trc_pfx << "Purge check:");
744 
745  ps.m_bytes_to_remove = 0;
746  if (conf.are_file_usage_limits_set())
747  {
748  ps.m_bytes_to_remove = get_file_usage_bytes_to_remove(ps, ps.m_estimated_writes_from_writeq, tl);
749  }
750  else
751  {
752  if (ps.m_disk_used > conf.m_diskUsageHWM)
753  {
754  TRACE_INT(tl, "Disk usage: " << ps.m_disk_used << " exceed highWatermark.");
756  }
757  }
758 
759  ps.m_space_based_purge = ps.m_bytes_to_remove ? 1 : 0;
760 
761  // Purge precheck -- check if age-based purge is required
762  // We ignore uvkeep time, it requires reading of cinfo files and it is enforced in File::Open() anyway.
763 
764  if (purge_cold_files && conf.is_age_based_purge_in_effect()) // || conf.is_uvkeep_purge_in_effect())
765  {
766  ps.m_age_based_purge = true;
767  }
768 
769  TRACE_INT(tl, "\tbytes_to_remove = " << ps.m_bytes_to_remove << " B");
770  TRACE_INT(tl, "\tspace_based_purge = " << ps.m_space_based_purge);
771  TRACE_INT(tl, "\tage_based_purge = " << ps.m_age_based_purge);
772 
773  bool periodic = Cache::GetInstance().GetPurgePin() ?
775 
776  if ( ! ps.m_space_based_purge && ! ps.m_age_based_purge && !periodic ) {
777  TRACE(Info, trc_pfx << "purge not required.");
779  return;
780  }
781  if (m_purge_task_active) {
782  TRACE(Warning, trc_pfx << "purge required but previous purge task is still active!");
783  return;
784  }
785 
786  TRACE(Info, trc_pfx << "scheduling purge task.");
787 
788  // At this point we have all the information: report, decide on action.
789  // There is still some missing infrastructure, especially as regards to purge-plugin:
790  // - at what point do we start bugging the pu-pin to start coughing up purge lists?
791  // - have a new parameter or just do it "one cycle before full"?
792  // - what if it doesn't -- when do we do the old-stlye scan & purge?
793  // - how do we do age-based purge and uvkeep purge?
794  // - they are really quite different -- and could run separately, registering
795  // files into a purge-candidate list. This has to be rechecked before the actual
796  // deletion -- eg, by comparing stat time of cinfo + doing the is-active / is-purge-protected.
797 
798  const DirState &root_ds = *m_fs_state.get_root();
799 
800  // Size from the tree itself; the usage counters can be off (#2808).
801  const int n_pshot_dirs = root_ds.count_dirs_to_level(9999);
802 
803  // Report a counters-vs-tree mismatch when it appears or changes.
804  const int n_calc_dirs = 1 + root_ds.m_here_usage.m_NDirectories
806  dprintf("purge dir count recursive=%d vs from_usage=%d\n", n_pshot_dirs, n_calc_dirs);
807  const int discrepancy = n_calc_dirs - n_pshot_dirs;
808  if (discrepancy != m_dir_count_discrepancy)
809  {
810  if (discrepancy != 0) {
811  TRACE(Error, trc_pfx << "directory count mismatch: usage counters " << n_calc_dirs
812  << ", tree " << n_pshot_dirs << ", previous difference " << m_dir_count_discrepancy);
813  } else {
814  TRACE(Info, trc_pfx << "directory count consistent again: " << n_pshot_dirs);
815  }
816  m_dir_count_discrepancy = discrepancy;
817  }
818 
819  ps.m_dir_vec.reserve(n_pshot_dirs);
820  ps.m_dir_vec.emplace_back( DirPurgeElement(root_ds, root_ds.m_here_usage, root_ds.m_recursive_subdir_usage, -1) );
821  fill_pshot_vec_children(root_ds, 0, ps.m_dir_vec, 9999);
822 
823  m_purge_task_active = true;
824 
825  struct PurgeDriverJob : public XrdJob
826  {
827  DataFsPurgeshot *m_purge_shot_ptr;
828 
829  PurgeDriverJob(DataFsPurgeshot *psp) :
830  XrdJob("XrdPfc::ResourceMonitor::PurgeDriver"),
831  m_purge_shot_ptr(psp)
832  {}
833 
834  void DoIt() override
835  {
836  Cache::ResMon().perform_purge_task(*m_purge_shot_ptr);
838 
839  delete m_purge_shot_ptr;
840  delete this;
841  }
842  };
843 
844  Cache::schedP->Schedule( new PurgeDriverJob(psp.release()) );
845 }
846 
847 namespace XrdPfc
848 {
850 }
851 
853 {
854  // BEWARE: Runs in a dedicated thread - is only to communicate back to the
855  // hear_beat() / data structs via the purge queues and condition variable.
856 
857  // const char *tpfx = "perform_purge_task ";
858 
859  {
861  m_purge_task_start = time(0);
862  }
863 
864  // For now, fall back to the old purge ... to be improved with:
865  // - new scan, following the DataFsPurgeshot;
866  // - usage of cinfo stat mtime for time of last access (touch already done at output);
867  // - use DirState* to report back purged files.
868  // Already changed to report back purged files --- but using the string / path variant.
869  OldStylePurgeDriver(ps); // In XrdPfcPurge.cc
870 }
871 
873 {
874  // Separated out so the purge_task can exit without post-checks.
875 
876  {
878  m_purge_task_end = time(0);
879  m_purge_task_complete = true;
881  }
883 }
884 
885 //==============================================================================
886 // Main thread function, do initial test, then enter heart_beat().
887 //==============================================================================
888 
890 {
891  // setup for in-scan -- this is called from initial setup.
892  MutexHolder _lck(m_dir_scan_mutex);
893  m_dir_scan_in_progress = true;
894 }
895 
897 {
898  const char *tpfx = "main_thread_function ";
899  {
900  time_t is_start = time(0);
901  m_fs_state.init_stat_reset_times(is_start);
902  TRACE(Info, tpfx << "Stating initial directory scan.");
903 
904  if ( ! perform_initial_scan()) {
905  TRACE(Error, tpfx << "Initial directory scan has failed. This is a terminal error, aborting.")
906  _exit(1);
907  }
908  // Reset of m_dir_scan_in_progress is done in perform_initial_scan()
909 
910  time_t is_duration = time(0) - is_start;
911  TRACE(Info, tpfx << "Initial directory scan complete, duration=" << is_duration <<"s");
912 
913  // run first process queues
914  int n_proc_is = process_queues();
915  TRACE(Info, tpfx << "First process_queues finished, n_records=" << n_proc_is);
916 
917  // shrink queues if scan time was longer than 30s.
918  if (is_duration > 30 || n_proc_is > 3000)
919  {
920  m_file_open_q.shrink_read_queue();
921  m_file_update_stats_q.shrink_read_queue();
922  m_file_close_q.shrink_read_queue();
923  m_file_purge_q1.shrink_read_queue();
924  m_file_purge_q2.shrink_read_queue();
925  m_file_purge_q3.shrink_read_queue();
926  }
927  }
928  heart_beat();
929 }
930 
931 //==============================================================================
932 // Old prototype from Cache / Purge, now to go into heart_beat() here, above.
933 //==============================================================================
934 
936 {
937  // static const char *trc_pfx = "ResourceMonitorHeartBeat() ";
938 
939  // Pause before initial run
940  sleep(1);
941 
942  // XXXX Setup initial / constant stats (total RAM, total disk, ???)
943 
946 
947  S.Lock();
948 
950 
952 
953  S.UnLock();
954 
955  // XXXX Schedule initial disk scan, time it!
956  //
957  // TRACE(Info, trc_pfx << "scheduling intial disk scan.");
958  // schedP->Schedule( new ScanAndPurgeJob("XrdPfc::ScanAndPurge") );
959  //
960  // bool scan_and_purge_running = true;
961 
962  // XXXX Could we really hold last-usage for all files in memory?
963 
964  // XXXX Think how to handle disk-full, scan/purge not finishing:
965  // - start dropping things out of write queue, but only when RAM gets near full;
966  // - monitoring this then becomes a high-priority job, inner loop with sleep of,
967  // say, 5 or 10 seconds.
968 
969  while (true)
970  {
971  time_t heartbeat_start = time(0);
972 
973  // TRACE(Info, trc_pfx << "HeartBeat starting ...");
974 
975  // if sumary monitoring configured, pupulate OucCacheStats:
976  S.Lock();
977 
978  // - available / used disk space (files usage calculated elsewhere (maybe))
979 
980  // - RAM usage
981  /* XXXX From Cache
982  { XrdSysMutexHelper lck(&m_RAM_mutex);
983  X.MemUsed = m_RAM_used;
984  X.MemWriteQ = m_RAM_write_queue;
985  }
986  */
987 
988  // - files opened / closed etc
989 
990  // do estimate of available space
991  S.UnLock();
992 
993  // if needed, schedule purge in a different thread.
994  // purge is:
995  // - deep scan + gather FSPurgeState
996  // - actual purge
997  //
998  // this thread can continue running and, if needed, stop writing to disk
999  // if purge is taking too long.
1000 
1001  // think how data is passed / synchronized between this and purge thread
1002 
1003  // !!!! think how stat collection is done and propgated upwards;
1004  // until now it was done once per purge-interval.
1005  // now stats will be added up more often, but purge will be done
1006  // only occasionally.
1007  // also, do we report cumulative values or deltas? cumulative should
1008  // be easier and consistent with summary data.
1009  // still, some are state - like disk usage, num of files.
1010 
1011  // Do we take care of directories that need to be newly added into DirState hierarchy?
1012  // I.e., when user creates new directories and these are covered by either full
1013  // spec or by root + depth declaration.
1014 
1015  int heartbeat_duration = time(0) - heartbeat_start;
1016 
1017  // TRACE(Info, trc_pfx << "HeartBeat finished, heartbeat_duration " << heartbeat_duration);
1018 
1019  // int sleep_time = m_fs_state..m_purgeInterval - heartbeat_duration;
1020  int sleep_time = 60 - heartbeat_duration;
1021  if (sleep_time > 0)
1022  {
1023  sleep(sleep_time);
1024  }
1025  }
1026 }
int DoIt(int argpnt, int argc, char **argv, bool singleshot)
Definition: XrdAccTest.cc:262
static void child()
#define TRACE_Debug
Definition: XrdCmsTrace.hh:37
#define TRACE_Info
#define XrdOssOK
Definition: XrdOss.hh:50
#define dprintf(...)
void Proto_ResourceMonitorHeartBeat()
#define TRACE_INT(act, x)
Definition: XrdPfcTrace.hh:52
bool Debug
@ Error
#define TRACE(act, x)
Definition: XrdTrace.hh:63
Definition: XrdJob.hh:43
virtual int Opendir(const char *path, XrdOucEnv &env)
Definition: XrdOss.hh:79
long long Total
Definition: XrdOssVS.hh:90
long long Free
Definition: XrdOssVS.hh:91
virtual XrdOssDF * newDir(const char *tident)=0
virtual int StatVS(XrdOssVSInfo *vsP, const char *sname=0, int updt=0)
Definition: XrdOss.cc:117
virtual int Unlink(const char *path, int Opts=0, XrdOucEnv *envP=0)=0
struct XrdOucCacheStats::CacheStats X
XrdOucCacheStats Statistics
Definition: XrdOucCache.hh:686
PurgePin * GetPurgePin() const
Definition: XrdPfc.hh:286
static const Configuration & Conf()
Definition: XrdPfc.cc:134
static ResourceMonitor & ResMon()
Definition: XrdPfc.cc:135
void ClearPurgeProtectedSet()
Definition: XrdPfc.cc:691
static Cache & GetInstance()
Singleton access.
Definition: XrdPfc.cc:132
XrdSysTrace * GetTrace() const
Definition: XrdPfc.hh:297
static XrdScheduler * schedP
Definition: XrdPfc.hh:304
long long WritesSinceLastCall()
Definition: XrdPfc.cc:324
void AddUp(const DirStats &s)
Definition: XrdPfcStats.hh:192
long long m_StBlocksRemoved
Definition: XrdPfcStats.hh:149
std::vector< std::string > m_current_dirs
std::function< void()> m_readdir_hook
XrdOucEnv & default_env()
void slurp_dir_ll(XrdOssDF &dh, int dir_level, const char *path, const char *trc_pfx)
bool begin_traversal(DirState *root, const char *root_path)
std::set< std::string > m_protected_top_dirs
bool cd_down(const std::string &dir_name)
std::map< std::string, FilePairStat > m_current_files
Status of cached file. Can be read from and written into a binary file.
Definition: XrdPfcInfo.hh:41
virtual bool CallPeriodically()
void CrossCheckIfScanIsInProgress(const std::string &lfn, XrdSysCondVar &cond)
void perform_purge_check(bool purge_cold_files, int tl)
void fill_sshot_vec_children(const DirState &parent_ds, int parent_idx, std::vector< DirStateElement > &vec, int max_depth)
void perform_purge_task(DataFsPurgeshot &ps)
void scan_dir_and_recurse(FsTraversal &fst)
void fill_pshot_vec_children(const DirState &parent_ds, int parent_idx, std::vector< DirPurgeElement > &vec, int max_depth)
void release(int i)
Clear the slot and return it for reuse. Caller must hold its mutex.
void Schedule(XrdJob *jp)
@ Warning
Definition: XrdPfc.hh:41
void OldStylePurgeDriver(DataFsPurgeshot &ps)
Definition: XrdPfcPurge.cc:99
void note_file_open(DirState *node, DirState *last_existing, bool existing_file)
std::function< int(const std::string &)> unlink_func
Contains parameters configurable from the xrootd config file.
Definition: XrdPfc.hh:64
long long m_RamAbsAvailable
available from configuration
Definition: XrdPfc.hh:109
long long m_diskTotalSpace
total disk space on configured partition or oss space
Definition: XrdPfc.hh:92
long long m_fileUsageMax
cache purge - files usage maximum
Definition: XrdPfc.hh:97
long long m_fileUsageBaseline
cache purge - files usage baseline
Definition: XrdPfc.hh:95
int m_dirStatsStoreDepth
maximum depth for statistics write out
Definition: XrdPfc.hh:106
long long m_diskUsageHWM
cache purge - disk usage high water mark
Definition: XrdPfc.hh:94
bool are_file_usage_limits_set() const
Definition: XrdPfc.hh:67
long long m_fileUsageNominal
cache purge - files usage nominal
Definition: XrdPfc.hh:96
int m_purgeAgeBasedPeriod
peform cold file / uvkeep purge every this many purge cycles
Definition: XrdPfc.hh:100
long long m_diskUsageLWM
cache purge - disk usage low water mark
Definition: XrdPfc.hh:93
int m_dirStatsInterval
time between resource monitor statistics dump in seconds
Definition: XrdPfc.hh:105
bool is_age_based_purge_in_effect() const
Definition: XrdPfc.hh:68
int m_purgeInterval
sleep interval between cache purges
Definition: XrdPfc.hh:98
bool is_dir_stat_reporting_on() const
Definition: XrdPfc.hh:70
std::vector< DirPurgeElement > m_dir_vec
void write_json_file(const std::string &fname, XrdOss &oss, bool include_preamble)
std::vector< DirStateElement > m_dir_states
void reset_stats(time_t last_update)
void dump_recursively(int max_depth) const
void init_stat_reset_times(time_t t)
void update_stats_and_usages(time_t last_update, bool purge_empty_dirs, unlink_func unlink_foo)
DirState * find_dirstate_for_lfn(const std::string &lfn, DirState **last_existing_dir=nullptr)
void reset_sshot_stats(time_t last_update)
DirUsage m_recursive_subdir_usage
int count_dirs_to_level(int max_depth) const
DirState * find_path(const std::string &path, int max_depth, bool parse_as_lfn, bool create_subdirs, DirState **last_existing_dir=nullptr)
void upward_propagate_initial_scan_usages()