1/*-------------------------------------------------------------------------
2 *
3 * ipc.c
4 * POSTGRES inter-process communication definitions.
5 *
6 * This file is misnamed, as it no longer has much of anything directly
7 * to do with IPC. The functionality here is concerned with managing
8 * exit-time cleanup for either a postmaster or a backend.
9 *
10 *
11 * Portions Copyright (c) 1996-2019, PostgreSQL Global Development Group
12 * Portions Copyright (c) 1994, Regents of the University of California
13 *
14 *
15 * IDENTIFICATION
16 * src/backend/storage/ipc/ipc.c
17 *
18 *-------------------------------------------------------------------------
19 */
20#include "postgres.h"
21
22#include <signal.h>
23#include <unistd.h>
24#include <sys/stat.h>
25
26#include "miscadmin.h"
27#ifdef PROFILE_PID_DIR
28#include "postmaster/autovacuum.h"
29#endif
30#include "storage/dsm.h"
31#include "storage/ipc.h"
32#include "tcop/tcopprot.h"
33
34
35/*
36 * This flag is set during proc_exit() to change ereport()'s behavior,
37 * so that an ereport() from an on_proc_exit routine cannot get us out
38 * of the exit procedure. We do NOT want to go back to the idle loop...
39 */
40bool proc_exit_inprogress = false;
41
42/*
43 * Set when shmem_exit() is in progress.
44 */
45bool shmem_exit_inprogress = false;
46
47/*
48 * This flag tracks whether we've called atexit() in the current process
49 * (or in the parent postmaster).
50 */
51static bool atexit_callback_setup = false;
52
53/* local functions */
54static void proc_exit_prepare(int code);
55
56
57/* ----------------------------------------------------------------
58 * exit() handling stuff
59 *
60 * These functions are in generally the same spirit as atexit(),
61 * but provide some additional features we need --- in particular,
62 * we want to register callbacks to invoke when we are disconnecting
63 * from a broken shared-memory context but not exiting the postmaster.
64 *
65 * Callback functions can take zero, one, or two args: the first passed
66 * arg is the integer exitcode, the second is the Datum supplied when
67 * the callback was registered.
68 * ----------------------------------------------------------------
69 */
70
71#define MAX_ON_EXITS 20
72
73struct ONEXIT
74{
75 pg_on_exit_callback function;
76 Datum arg;
77};
78
79static struct ONEXIT on_proc_exit_list[MAX_ON_EXITS];
80static struct ONEXIT on_shmem_exit_list[MAX_ON_EXITS];
81static struct ONEXIT before_shmem_exit_list[MAX_ON_EXITS];
82
83static int on_proc_exit_index,
84 on_shmem_exit_index,
85 before_shmem_exit_index;
86
87
88/* ----------------------------------------------------------------
89 * proc_exit
90 *
91 * this function calls all the callbacks registered
92 * for it (to free resources) and then calls exit.
93 *
94 * This should be the only function to call exit().
95 * -cim 2/6/90
96 *
97 * Unfortunately, we can't really guarantee that add-on code
98 * obeys the rule of not calling exit() directly. So, while
99 * this is the preferred way out of the system, we also register
100 * an atexit callback that will make sure cleanup happens.
101 * ----------------------------------------------------------------
102 */
103void
104proc_exit(int code)
105{
106 /* Clean up everything that must be cleaned up */
107 proc_exit_prepare(code);
108
109#ifdef PROFILE_PID_DIR
110 {
111 /*
112 * If we are profiling ourself then gprof's mcleanup() is about to
113 * write out a profile to ./gmon.out. Since mcleanup() always uses a
114 * fixed file name, each backend will overwrite earlier profiles. To
115 * fix that, we create a separate subdirectory for each backend
116 * (./gprof/pid) and 'cd' to that subdirectory before we exit() - that
117 * forces mcleanup() to write each profile into its own directory. We
118 * end up with something like: $PGDATA/gprof/8829/gmon.out
119 * $PGDATA/gprof/8845/gmon.out ...
120 *
121 * To avoid undesirable disk space bloat, autovacuum workers are
122 * discriminated against: all their gmon.out files go into the same
123 * subdirectory. Without this, an installation that is "just sitting
124 * there" nonetheless eats megabytes of disk space every few seconds.
125 *
126 * Note that we do this here instead of in an on_proc_exit() callback
127 * because we want to ensure that this code executes last - we don't
128 * want to interfere with any other on_proc_exit() callback. For the
129 * same reason, we do not include it in proc_exit_prepare ... so if
130 * you are exiting in the "wrong way" you won't drop your profile in a
131 * nice place.
132 */
133 char gprofDirName[32];
134
135 if (IsAutoVacuumWorkerProcess())
136 snprintf(gprofDirName, 32, "gprof/avworker");
137 else
138 snprintf(gprofDirName, 32, "gprof/%d", (int) getpid());
139
140 /*
141 * Use mkdir() instead of MakePGDirectory() since we aren't making a
142 * PG directory here.
143 */
144 mkdir("gprof", S_IRWXU | S_IRWXG | S_IRWXO);
145 mkdir(gprofDirName, S_IRWXU | S_IRWXG | S_IRWXO);
146 chdir(gprofDirName);
147 }
148#endif
149
150 elog(DEBUG3, "exit(%d)", code);
151
152 exit(code);
153}
154
155/*
156 * Code shared between proc_exit and the atexit handler. Note that in
157 * normal exit through proc_exit, this will actually be called twice ...
158 * but the second call will have nothing to do.
159 */
160static void
161proc_exit_prepare(int code)
162{
163 /*
164 * Once we set this flag, we are committed to exit. Any ereport() will
165 * NOT send control back to the main loop, but right back here.
166 */
167 proc_exit_inprogress = true;
168
169 /*
170 * Forget any pending cancel or die requests; we're doing our best to
171 * close up shop already. Note that the signal handlers will not set
172 * these flags again, now that proc_exit_inprogress is set.
173 */
174 InterruptPending = false;
175 ProcDiePending = false;
176 QueryCancelPending = false;
177 InterruptHoldoffCount = 1;
178 CritSectionCount = 0;
179
180 /*
181 * Also clear the error context stack, to prevent error callbacks from
182 * being invoked by any elog/ereport calls made during proc_exit. Whatever
183 * context they might want to offer is probably not relevant, and in any
184 * case they are likely to fail outright after we've done things like
185 * aborting any open transaction. (In normal exit scenarios the context
186 * stack should be empty anyway, but it might not be in the case of
187 * elog(FATAL) for example.)
188 */
189 error_context_stack = NULL;
190 /* For the same reason, reset debug_query_string before it's clobbered */
191 debug_query_string = NULL;
192
193 /* do our shared memory exits first */
194 shmem_exit(code);
195
196 elog(DEBUG3, "proc_exit(%d): %d callbacks to make",
197 code, on_proc_exit_index);
198
199 /*
200 * call all the registered callbacks.
201 *
202 * Note that since we decrement on_proc_exit_index each time, if a
203 * callback calls ereport(ERROR) or ereport(FATAL) then it won't be
204 * invoked again when control comes back here (nor will the
205 * previously-completed callbacks). So, an infinite loop should not be
206 * possible.
207 */
208 while (--on_proc_exit_index >= 0)
209 on_proc_exit_list[on_proc_exit_index].function(code,
210 on_proc_exit_list[on_proc_exit_index].arg);
211
212 on_proc_exit_index = 0;
213}
214
215/* ------------------
216 * Run all of the on_shmem_exit routines --- but don't actually exit.
217 * This is used by the postmaster to re-initialize shared memory and
218 * semaphores after a backend dies horribly. As with proc_exit(), we
219 * remove each callback from the list before calling it, to avoid
220 * infinite loop in case of error.
221 * ------------------
222 */
223void
224shmem_exit(int code)
225{
226 shmem_exit_inprogress = true;
227
228 /*
229 * Call before_shmem_exit callbacks.
230 *
231 * These should be things that need most of the system to still be up and
232 * working, such as cleanup of temp relations, which requires catalog
233 * access; or things that need to be completed because later cleanup steps
234 * depend on them, such as releasing lwlocks.
235 */
236 elog(DEBUG3, "shmem_exit(%d): %d before_shmem_exit callbacks to make",
237 code, before_shmem_exit_index);
238 while (--before_shmem_exit_index >= 0)
239 before_shmem_exit_list[before_shmem_exit_index].function(code,
240 before_shmem_exit_list[before_shmem_exit_index].arg);
241 before_shmem_exit_index = 0;
242
243 /*
244 * Call dynamic shared memory callbacks.
245 *
246 * These serve the same purpose as late callbacks, but for dynamic shared
247 * memory segments rather than the main shared memory segment.
248 * dsm_backend_shutdown() has the same kind of progressive logic we use
249 * for the main shared memory segment; namely, it unregisters each
250 * callback before invoking it, so that we don't get stuck in an infinite
251 * loop if one of those callbacks itself throws an ERROR or FATAL.
252 *
253 * Note that explicitly calling this function here is quite different from
254 * registering it as an on_shmem_exit callback for precisely this reason:
255 * if one dynamic shared memory callback errors out, the remaining
256 * callbacks will still be invoked. Thus, hard-coding this call puts it
257 * equal footing with callbacks for the main shared memory segment.
258 */
259 dsm_backend_shutdown();
260
261 /*
262 * Call on_shmem_exit callbacks.
263 *
264 * These are generally releasing low-level shared memory resources. In
265 * some cases, this is a backstop against the possibility that the early
266 * callbacks might themselves fail, leading to re-entry to this routine;
267 * in other cases, it's cleanup that only happens at process exit.
268 */
269 elog(DEBUG3, "shmem_exit(%d): %d on_shmem_exit callbacks to make",
270 code, on_shmem_exit_index);
271 while (--on_shmem_exit_index >= 0)
272 on_shmem_exit_list[on_shmem_exit_index].function(code,
273 on_shmem_exit_list[on_shmem_exit_index].arg);
274 on_shmem_exit_index = 0;
275
276 shmem_exit_inprogress = false;
277}
278
279/* ----------------------------------------------------------------
280 * atexit_callback
281 *
282 * Backstop to ensure that direct calls of exit() don't mess us up.
283 *
284 * Somebody who was being really uncooperative could call _exit(),
285 * but for that case we have a "dead man switch" that will make the
286 * postmaster treat it as a crash --- see pmsignal.c.
287 * ----------------------------------------------------------------
288 */
289static void
290atexit_callback(void)
291{
292 /* Clean up everything that must be cleaned up */
293 /* ... too bad we don't know the real exit code ... */
294 proc_exit_prepare(-1);
295}
296
297/* ----------------------------------------------------------------
298 * on_proc_exit
299 *
300 * this function adds a callback function to the list of
301 * functions invoked by proc_exit(). -cim 2/6/90
302 * ----------------------------------------------------------------
303 */
304void
305on_proc_exit(pg_on_exit_callback function, Datum arg)
306{
307 if (on_proc_exit_index >= MAX_ON_EXITS)
308 ereport(FATAL,
309 (errcode(ERRCODE_PROGRAM_LIMIT_EXCEEDED),
310 errmsg_internal("out of on_proc_exit slots")));
311
312 on_proc_exit_list[on_proc_exit_index].function = function;
313 on_proc_exit_list[on_proc_exit_index].arg = arg;
314
315 ++on_proc_exit_index;
316
317 if (!atexit_callback_setup)
318 {
319 atexit(atexit_callback);
320 atexit_callback_setup = true;
321 }
322}
323
324/* ----------------------------------------------------------------
325 * before_shmem_exit
326 *
327 * Register early callback to perform user-level cleanup,
328 * e.g. transaction abort, before we begin shutting down
329 * low-level subsystems.
330 * ----------------------------------------------------------------
331 */
332void
333before_shmem_exit(pg_on_exit_callback function, Datum arg)
334{
335 if (before_shmem_exit_index >= MAX_ON_EXITS)
336 ereport(FATAL,
337 (errcode(ERRCODE_PROGRAM_LIMIT_EXCEEDED),
338 errmsg_internal("out of before_shmem_exit slots")));
339
340 before_shmem_exit_list[before_shmem_exit_index].function = function;
341 before_shmem_exit_list[before_shmem_exit_index].arg = arg;
342
343 ++before_shmem_exit_index;
344
345 if (!atexit_callback_setup)
346 {
347 atexit(atexit_callback);
348 atexit_callback_setup = true;
349 }
350}
351
352/* ----------------------------------------------------------------
353 * on_shmem_exit
354 *
355 * Register ordinary callback to perform low-level shutdown
356 * (e.g. releasing our PGPROC); run after before_shmem_exit
357 * callbacks and before on_proc_exit callbacks.
358 * ----------------------------------------------------------------
359 */
360void
361on_shmem_exit(pg_on_exit_callback function, Datum arg)
362{
363 if (on_shmem_exit_index >= MAX_ON_EXITS)
364 ereport(FATAL,
365 (errcode(ERRCODE_PROGRAM_LIMIT_EXCEEDED),
366 errmsg_internal("out of on_shmem_exit slots")));
367
368 on_shmem_exit_list[on_shmem_exit_index].function = function;
369 on_shmem_exit_list[on_shmem_exit_index].arg = arg;
370
371 ++on_shmem_exit_index;
372
373 if (!atexit_callback_setup)
374 {
375 atexit(atexit_callback);
376 atexit_callback_setup = true;
377 }
378}
379
380/* ----------------------------------------------------------------
381 * cancel_before_shmem_exit
382 *
383 * this function removes a previously-registered before_shmem_exit
384 * callback. For simplicity, only the latest entry can be
385 * removed. (We could work harder but there is no need for
386 * current uses.)
387 * ----------------------------------------------------------------
388 */
389void
390cancel_before_shmem_exit(pg_on_exit_callback function, Datum arg)
391{
392 if (before_shmem_exit_index > 0 &&
393 before_shmem_exit_list[before_shmem_exit_index - 1].function
394 == function &&
395 before_shmem_exit_list[before_shmem_exit_index - 1].arg == arg)
396 --before_shmem_exit_index;
397}
398
399/* ----------------------------------------------------------------
400 * on_exit_reset
401 *
402 * this function clears all on_proc_exit() and on_shmem_exit()
403 * registered functions. This is used just after forking a backend,
404 * so that the backend doesn't believe it should call the postmaster's
405 * on-exit routines when it exits...
406 * ----------------------------------------------------------------
407 */
408void
409on_exit_reset(void)
410{
411 before_shmem_exit_index = 0;
412 on_shmem_exit_index = 0;
413 on_proc_exit_index = 0;
414 reset_on_dsm_detach();
415}
416