1 | /* |
2 | * This Source Code Form is subject to the terms of the Mozilla Public |
3 | * License, v. 2.0. If a copy of the MPL was not distributed with this |
4 | * file, You can obtain one at http://mozilla.org/MPL/2.0/. |
5 | * |
6 | * Copyright 1997 - July 2008 CWI, August 2008 - 2019 MonetDB B.V. |
7 | */ |
8 | |
9 | /* |
10 | * @a Niels Nes, Peter Boncz |
11 | * @* System Independent Layer |
12 | * |
13 | * GDK is built on Posix. Exceptions are made for memory mapped files |
14 | * and anonymous virtual memory, for which somewhat higher-level |
15 | * functions are defined here. Most of this file concerns itself with |
16 | * emulation of Posix functionality on the WIN32 native platform. |
17 | */ |
18 | #include "monetdb_config.h" |
19 | #include "gdk.h" /* includes gdk_posix.h */ |
20 | #include "gdk_private.h" |
21 | #include "mutils.h" |
22 | #include <unistd.h> |
23 | #include <string.h> /* strncpy */ |
24 | |
25 | #ifdef HAVE_FCNTL_H |
26 | # include <fcntl.h> |
27 | #endif |
28 | #ifdef HAVE_PROCFS_H |
29 | # include <procfs.h> |
30 | #endif |
31 | #ifdef HAVE_MACH_TASK_H |
32 | # include <mach/task.h> |
33 | #endif |
34 | #ifdef HAVE_MACH_MACH_INIT_H |
35 | # include <mach/mach_init.h> |
36 | #endif |
37 | #if defined(HAVE_KVM_H) |
38 | # include <kvm.h> |
39 | # include <sys/param.h> |
40 | # include <sys/sysctl.h> |
41 | # include <sys/user.h> |
42 | #endif |
43 | |
44 | #ifdef NDEBUG |
45 | #ifndef NVALGRIND |
46 | #define NVALGRIND NDEBUG |
47 | #endif |
48 | #endif |
49 | |
50 | #if defined(__GNUC__) && defined(HAVE_VALGRIND) |
51 | #include <valgrind.h> |
52 | #else |
53 | #define VALGRIND_MALLOCLIKE_BLOCK(addr, sizeB, rzB, is_zeroed) |
54 | #define VALGRIND_FREELIKE_BLOCK(addr, rzB) |
55 | #define VALGRIND_RESIZEINPLACE_BLOCK(addr, oldSizeB, newSizeB, rzB) |
56 | #endif |
57 | |
58 | #ifndef MAP_NORESERVE |
59 | # define MAP_NORESERVE MAP_PRIVATE |
60 | #endif |
61 | #if defined(MAP_ANON) && !defined(MAP_ANONYMOUS) |
62 | #define MAP_ANONYMOUS MAP_ANON |
63 | #endif |
64 | |
65 | #define MMAP_ADVISE 7 |
66 | #define MMAP_WRITABLE (MMAP_WRITE|MMAP_COPY) |
67 | |
68 | #ifndef O_CLOEXEC |
69 | #define O_CLOEXEC 0 |
70 | #endif |
71 | |
72 | /* DDALERT: AIX4.X 64bits needs HAVE_SETENV==0 due to a AIX bug, but |
73 | * it probably isn't detected so by configure */ |
74 | |
75 | #ifndef HAVE_SETENV |
76 | int |
77 | setenv(const char *name, const char *value, int overwrite) |
78 | { |
79 | int ret = 0; |
80 | |
81 | if (overwrite || getenv(name) == NULL) { |
82 | char *p = GDKmalloc(2 + strlen(name) + strlen(value)); |
83 | |
84 | if (p == NULL) |
85 | return -1; |
86 | strcpy(p, name); |
87 | strcat(p, "=" ); |
88 | strcat(p, value); |
89 | ret = putenv(p); |
90 | /* GDKfree(p); LEAK INSERTED DUE TO SOME WEIRD CRASHES */ |
91 | } |
92 | return ret; |
93 | } |
94 | #endif |
95 | |
96 | /* Crude VM buffer management that keep a list of all memory mapped |
97 | * regions. |
98 | * |
99 | * a.k.a. "helping stupid VM implementations that ignore VM advice" |
100 | * |
101 | * The main goal is to be able to tell the OS to please stop buffering |
102 | * all memory mapped pages when under pressure. A major problem is |
103 | * materialization of large results in newly created memory mapped |
104 | * files. Operating systems tend to cache all dirty pages, such that |
105 | * when memory is out, all pages are dirty and cannot be unloaded |
106 | * quickly. The VM panic occurs and comatose OS states may be |
107 | * observed. This is in spite of our use of |
108 | * madvise(MADV_SEQUENTIAL). That is; we would want that the OS drops |
109 | * pages after we've passed them. That does not happen; pages are |
110 | * retained and pollute the buffer cache. |
111 | * |
112 | * Regrettably, at this level, we don't know anything about how Monet |
113 | * is using the mmapped regions. Monet code is totally oblivious of |
114 | * any I/O; that's why it is so easy to create CPU efficient code in |
115 | * Monet. |
116 | * |
117 | * The current solution focuses on large writable maps. These often |
118 | * represent newly created BATs, that are the result of some (running) |
119 | * operator. We assume two things here: |
120 | * - the BAT is created in sequential fashion (always almost true) |
121 | * - afterwards, this BAT is used in sequential fashion (often true) |
122 | * |
123 | * A VMtrim thread keeps an eye on the RSS (memory pressure) and large |
124 | * writable memory maps. If RSS approaches mem_maxsize(), it starts to |
125 | * *worry*, and starts to write dirty data from these writable maps to |
126 | * disk in 128MB tiles. So, if memory pressure rises further in the |
127 | * near future, the OS has some option to release memory pages cheaply |
128 | * (i.e. without needing I/O). This is also done explicitly by the |
129 | * VM-thread: when RSS exceeds mem_maxsize() is explicitly asks the OS |
130 | * to release pages. The reason is that Linux is not smart enough to |
131 | * do even this. Anyway.. |
132 | * |
133 | * The way to free pages explicitly in Linux is to call |
134 | * posix_fadvise(..,MADV_DONTNEED). Particularly, |
135 | * posix_madvise(..,POSIX_MADV_DONTNEED) which is supported and |
136 | * documented doesn't work on Linux. But we do both posix_madvise and |
137 | * posix_fadvise, so on other unix systems that don't support |
138 | * posix_fadvise, posix_madvise still might work. On Windows, to our |
139 | * knowledge, there is no way to tell it stop buffering a memory |
140 | * mapped region. msync (FlushViewOfFile) does work, though. So let's |
141 | * hope the VM paging algorithm behaves better than Linux which just |
142 | * runs off the cliff and if MonetDB does not prevent RSS from being |
143 | * too high, enters coma. |
144 | * |
145 | * We will only be able to sensibly test this on Windows64. On |
146 | * Windows32, mmap sizes do not significantly exceed RAM sizes so |
147 | * MonetDB swapping actually will not happen (of course, you've got |
148 | * this nasty problem of VM fragemntation and failing mmaps instead). |
149 | * |
150 | * In principle, page tiles are saved sequentially, and behind it, but |
151 | * never overtaking it, is an "unload-cursor" that frees the pages if |
152 | * that is needed to keep RSS down. There is a tweak in the |
153 | * algorithm, that re-sets the unload-cursor if it seems that all |
154 | * tiles to the end have been saved (whether a tile is actually saved |
155 | * is determined by timing the sync action). This means that the |
156 | * producing operator is ready creating the BAT, and we assume it is |
157 | * going to be used sequentially afterwards. In that case, we should |
158 | * start unloading right after the 'read-cursor', that is, from the |
159 | * start. |
160 | * |
161 | * EXAMPLE |
162 | * D = dirty tile |
163 | * s = saved tile (i.e. clean) |
164 | * u = unloaded tile |
165 | * L = tile that is being loaded |
166 | * |
167 | * +--> operator produces BAT |
168 | * (1) DDDDDD|......................................| end of reserved mmap |
169 | * ____|RSS |
170 | * | |
171 | * | at 3/4 of RSS consumed we start to worry |
172 | * +--> operator produces BAT |
173 | * (2) DDDDDDDDDDDDDDDD|............................| |
174 | * s<----------------------------- VM backwards save thread |
175 | * | |
176 | * + first tile of which saving costs anything |
177 | * |
178 | * +--> operator produces BAT |
179 | * (3) DDDDDDDDDDDDDDDss|D|.........................| |
180 | * VM-thread save ->| |
181 | * |
182 | * When the RSS target is exceeded, we start unloading tiles.. |
183 | * |
184 | * +--> VM-thread unload starts at *second* 's' |
185 | * | |
186 | * | +--> operator produces BAT |
187 | * (4) DDDDDDDDDDDDDDDsus|DD|........................| |
188 | * VM-thread save -->| | RSS = Full! |
189 | * |
190 | * +-- 0 => save costs nothing!! |
191 | * VM-thread save ------------->| assume bat complete |
192 | * (5) DDDDDDDDDDDDDDDsuuuuuuuuussss0................| |
193 | * |<-------- re-set unload cursor |
194 | * +--- first tile was not unloaded. |
195 | * |
196 | * later.. some other operator sequentially reads the bat |
197 | * first part is 'D', that is, nicely cached. |
198 | * |
199 | * ---read------->| |
200 | * (6) DDDDDDDDDDDDDDDsuuuuuuuuussss0................| |
201 | * |
202 | * now we're hitting the unloaded region. the query becomes |
203 | * I/O read bound here (typically 20% CPU utilization). |
204 | * |
205 | * ---read-------->| |
206 | * (7) DDDDDDDDDDDDDDDuLuuuuuuuussss0................| |
207 | * / \ |
208 | * unload cursor load cursor |
209 | * |
210 | * ---read---------------->| |
211 | * (8) DDDDDDDDDDDDDDDuuuuuuuuuLssss0................| |
212 | * / \ |
213 | * unload cursor load cursor |
214 | * |
215 | * ---read--------------------->| done |
216 | * (9) DDDDDDDDDDDDDDDuuuuuuuuuLssss0................| |
217 | * **** |
218 | * last part still cached |
219 | * |
220 | * note: if we would not have re-setted the unload cursor (5) |
221 | * the last part would have been lost due to continuing |
222 | * RSS pressure from the 'L' read-cursor. |
223 | * |
224 | * If multiple write-mmaps exist, we do unload-tile and save-tile |
225 | * selection on a round-robin basis among them. |
226 | * |
227 | * Of course, this is a simple solution for simple cases only. |
228 | * (a) if the bat is produced too fast, (or your disk is too slow) |
229 | * RSS will exceeds its limit and Linux will go into swapping. |
230 | * (b) if your data is not produced and read sequentially. |
231 | * Examples are sorting or clustering on huge datasets. |
232 | * (c) if RSS pressure is due to large read-maps, rather than |
233 | * intermediate results. |
234 | * |
235 | * Two crude suggestions: |
236 | * - If we are under RSS pressure without unloadable tiles and with |
237 | * savable tiles, we should consider suspending *all* other threads |
238 | * until we manage to unload a tile. |
239 | * - if there are no savable tiles (or in case of read-only maps) |
240 | * we could resort to saving and unloading random tiles. |
241 | * |
242 | * To do better, our BAT algorithms should provide even more detailed |
243 | * advice on their access patterns, which may even consist of pointers |
244 | * to the cursors (i.e. pointers to b->batBuns->free or the cursors |
245 | * in radix-cluster), which an enhanced version of this thread might |
246 | * take into account. |
247 | * |
248 | * [Kersten] The memory map table should be aligned to the number of |
249 | * mapped files. In more recent applications, such as the SkyServer |
250 | * this may be around 2000 BATs easily. |
251 | */ |
252 | |
253 | #ifdef HAVE_PTHREAD_H |
254 | /* pthread.h on Windows includes config.h if HAVE_CONFIG_H is set */ |
255 | #undef HAVE_CONFIG_H |
256 | #include <sched.h> |
257 | #include <pthread.h> |
258 | #endif |
259 | #ifdef HAVE_SEMAPHORE_H |
260 | #include <semaphore.h> |
261 | #endif |
262 | |
263 | #ifndef NATIVE_WIN32 |
264 | #ifdef HAVE_POSIX_FADVISE |
265 | #ifdef HAVE_UNAME |
266 | #include <sys/utsname.h> |
267 | #endif |
268 | #endif |
269 | |
270 | void |
271 | MT_init_posix(void) |
272 | { |
273 | } |
274 | |
275 | /* return RSS in bytes */ |
276 | size_t |
277 | (void) |
278 | { |
279 | #if defined(HAVE_PROCFS_H) && defined(__sun__) |
280 | /* retrieve RSS the Solaris way (2.6+) */ |
281 | int fd; |
282 | psinfo_t psbuff; |
283 | |
284 | fd = open("/proc/self/psinfo" , O_RDONLY | O_CLOEXEC); |
285 | if (fd >= 0) { |
286 | if (read(fd, &psbuff, sizeof(psbuff)) == sizeof(psbuff)) { |
287 | close(fd); |
288 | return psbuff.pr_rssize * 1024; |
289 | } |
290 | close(fd); |
291 | } |
292 | #elif defined(HAVE_TASK_INFO) |
293 | /* Darwin/MACH call for process' RSS */ |
294 | task_t task = mach_task_self(); |
295 | struct task_basic_info_64 t_info; |
296 | mach_msg_type_number_t t_info_count = TASK_BASIC_INFO_64_COUNT; |
297 | |
298 | if (task_info(task, TASK_BASIC_INFO_64, (task_info_t)&t_info, &t_info_count) != KERN_INVALID_POLICY) |
299 | return t_info.resident_size; /* bytes */ |
300 | #elif defined(HAVE_KVM_H) |
301 | /* get RSS on FreeBSD and NetBSD */ |
302 | struct kinfo_proc *ki; |
303 | int ski = 1; |
304 | kvm_t *kd; |
305 | size_t rss = 0; |
306 | |
307 | kd = kvm_open(NULL, "/dev/null" , NULL, O_RDONLY, "kvm_open" ); |
308 | if (kd != NULL) { |
309 | ki = kvm_getprocs(kd, KERN_PROC_PID, getpid(), &ski); |
310 | if (ki != NULL) { |
311 | #ifdef __NetBSD__ /* should we use configure for this? */ |
312 | /* see bug 3217 */ |
313 | rss = ki->kp_eproc.e_vm.vm_rssize; |
314 | #else |
315 | rss = ki->ki_rssize; |
316 | #endif |
317 | kvm_close(kd); |
318 | |
319 | return rss * MT_pagesize(); |
320 | } else { |
321 | kvm_close(kd); |
322 | } |
323 | } |
324 | #elif defined(__linux__) |
325 | /* get RSS on Linux */ |
326 | int fd; |
327 | |
328 | fd = open("/proc/self/stat" , O_RDONLY | O_CLOEXEC); |
329 | if (fd >= 0) { |
330 | char buf[1024], *r = buf; |
331 | ssize_t i, sz = read(fd, buf, 1024); |
332 | |
333 | close(fd); |
334 | if (sz > 0) { |
335 | for (i = 0; i < 23; i++) { |
336 | while (*r && (*r == ' ' || *r == '\t')) |
337 | r++; |
338 | while (*r && (*r != ' ' && *r != '\t')) |
339 | r++; |
340 | } |
341 | while (*r && (*r == ' ' || *r == '\t')) |
342 | r++; |
343 | return ((size_t) atol(r)) * MT_pagesize(); |
344 | } |
345 | } |
346 | #endif |
347 | return 0; |
348 | } |
349 | |
350 | void * |
351 | MT_mmap(const char *path, int mode, size_t len) |
352 | { |
353 | int fd; |
354 | void *ret; |
355 | |
356 | fd = open(path, O_CREAT | ((mode & MMAP_WRITE) ? O_RDWR : O_RDONLY) | O_CLOEXEC, MONETDB_MODE); |
357 | if (fd < 0) { |
358 | GDKsyserror("MT_mmap: open %s failed\n" , path); |
359 | return MAP_FAILED; |
360 | } |
361 | ret = mmap(NULL, |
362 | len, |
363 | ((mode & MMAP_WRITABLE) ? PROT_WRITE : 0) | PROT_READ, |
364 | (mode & MMAP_COPY) ? (MAP_PRIVATE | MAP_NORESERVE) : MAP_SHARED, |
365 | fd, |
366 | 0); |
367 | if (ret == MAP_FAILED) { |
368 | GDKsyserror("MT_mmap: mmap(%s,%zu) failed\n" , path, len); |
369 | ret = NULL; |
370 | } |
371 | close(fd); |
372 | VALGRIND_MALLOCLIKE_BLOCK(ret, len, 0, 1); |
373 | return ret; |
374 | } |
375 | |
376 | int |
377 | MT_munmap(void *p, size_t len) |
378 | { |
379 | int ret = munmap(p, len); |
380 | |
381 | if (ret < 0) |
382 | GDKsyserror("MT_munmap: munmap(%p,%zu) failed\n" , |
383 | p, len); |
384 | VALGRIND_FREELIKE_BLOCK(p, 0); |
385 | #ifdef MMAP_DEBUG |
386 | fprintf(stderr, "#munmap(%p,%zu) = %d\n" , p, len, ret); |
387 | #endif |
388 | return ret; |
389 | } |
390 | |
391 | /* expand or shrink a memory map (ala realloc). |
392 | * the address returned may be different from the address going in. |
393 | * in case of failure, the old address is still mapped and NULL is returned. |
394 | */ |
395 | void * |
396 | MT_mremap(const char *path, int mode, void *old_address, size_t old_size, size_t *new_size) |
397 | { |
398 | void *p; |
399 | int fd = -1; |
400 | int flags = mode & MMAP_COPY ? MAP_PRIVATE : MAP_SHARED; |
401 | int prot = PROT_WRITE | PROT_READ; |
402 | |
403 | /* round up to multiple of page size */ |
404 | *new_size = (*new_size + GDK_mmap_pagesize - 1) & ~(GDK_mmap_pagesize - 1); |
405 | |
406 | /* doesn't make sense for us to extend read-only memory map */ |
407 | assert(mode & MMAP_WRITABLE); |
408 | |
409 | if (*new_size < old_size) { |
410 | #ifndef STATIC_CODE_ANALYSIS /* hide this from static code analyzer */ |
411 | /* shrink */ |
412 | VALGRIND_RESIZEINPLACE_BLOCK(old_address, old_size, *new_size, 0); |
413 | if (munmap((char *) old_address + *new_size, |
414 | old_size - *new_size) < 0) { |
415 | GDKsyserror("MT_mremap: munmap(%p,%zu) failed\n" , |
416 | ((char *) old_address + *new_size), |
417 | old_size - *new_size); |
418 | fprintf(stderr, "= %s:%d: MT_mremap(%s,%p,%zu,%zu): munmap() failed\n" , __FILE__, __LINE__, path?path:"NULL" , old_address, old_size, *new_size); |
419 | /* even though the system call failed, we |
420 | * don't need to propagate the error up: the |
421 | * address should still work in the same way |
422 | * as it did before */ |
423 | return old_address; |
424 | } |
425 | if (path && truncate(path, *new_size) < 0) |
426 | fprintf(stderr, "#MT_mremap(%s): truncate failed\n" , path); |
427 | #ifdef MMAP_DEBUG |
428 | fprintf(stderr, "MT_mremap(%s,%p,%zu,%zu) -> shrinking\n" , path?path:"NULL" , old_address, old_size, *new_size); |
429 | #endif |
430 | #endif /* !STATIC_CODE_ANALYSIS */ |
431 | return old_address; |
432 | } |
433 | if (*new_size == old_size) { |
434 | /* do nothing */ |
435 | #ifdef MMAP_DEBUG |
436 | fprintf(stderr, "MT_mremap(%s,%p,%zu,%zu) -> unchanged\n" , path?path:"NULL" , old_address, old_size, *new_size); |
437 | #endif |
438 | return old_address; |
439 | } |
440 | |
441 | if (!(mode & MMAP_COPY) && path != NULL) { |
442 | /* "normal" memory map */ |
443 | |
444 | if ((fd = open(path, O_RDWR | O_CLOEXEC)) < 0) { |
445 | GDKsyserror("MT_mremap: open(%s) failed\n" , path); |
446 | fprintf(stderr, "= %s:%d: MT_mremap(%s,%p,%zu,%zu): open() failed\n" , __FILE__, __LINE__, path, old_address, old_size, *new_size); |
447 | return NULL; |
448 | } |
449 | if (GDKextendf(fd, *new_size, path) != GDK_SUCCEED) { |
450 | close(fd); |
451 | fprintf(stderr, "= %s:%d: MT_mremap(%s,%p,%zu,%zu): GDKextendf() failed\n" , __FILE__, __LINE__, path, old_address, old_size, *new_size); |
452 | return NULL; |
453 | } |
454 | #ifdef HAVE_MREMAP |
455 | /* on Linux it's easy */ |
456 | p = mremap(old_address, old_size, *new_size, MREMAP_MAYMOVE); |
457 | if (p == MAP_FAILED) |
458 | GDKsyserror("MT_mremap: mremap(%p,%zu,%zu) failed\n" , |
459 | old_address, old_size, |
460 | *new_size); |
461 | #ifdef HAVE_VALGRIND |
462 | if (p == old_address) { |
463 | VALGRIND_RESIZEINPLACE_BLOCK(old_address, old_size, *new_size, 0); |
464 | } else { |
465 | VALGRIND_FREELIKE_BLOCK(old_address, 0); |
466 | VALGRIND_MALLOCLIKE_BLOCK(p, *new_size, 0, 1); |
467 | } |
468 | #endif |
469 | #else |
470 | /* try to map extension at end of current map */ |
471 | p = mmap((char *) old_address + old_size, *new_size - old_size, |
472 | prot, flags, fd, old_size); |
473 | /* if it failed, there is no point trying a full mmap: |
474 | * that too won't fit */ |
475 | if (p != MAP_FAILED) { |
476 | if (p == (char *) old_address + old_size) { |
477 | /* we got the requested address, make |
478 | * sure we return the correct (old) |
479 | * address */ |
480 | VALGRIND_RESIZEINPLACE_BLOCK(old_address, old_size, *new_size, 0); |
481 | p = old_address; |
482 | } else { |
483 | /* we got some other address: discard |
484 | * it and make full mmap */ |
485 | munmap(p, *new_size - old_size); |
486 | #ifdef NO_MMAP_ALIASING |
487 | msync(old_address, old_size, MS_SYNC); |
488 | #endif |
489 | /* first create full mmap, then, if |
490 | * successful, remove old mmap */ |
491 | p = mmap(NULL, *new_size, prot, flags, fd, 0); |
492 | if (p != MAP_FAILED) { |
493 | VALGRIND_MALLOCLIKE_BLOCK(p, *new_size, 0, 1); |
494 | munmap(old_address, old_size); |
495 | VALGRIND_FREELIKE_BLOCK(old_address, 0); |
496 | } |
497 | } |
498 | } |
499 | if (p == MAP_FAILED) |
500 | GDKsyserror("MT_mremap: mmap failed\n" ); |
501 | #endif /* HAVE_MREMAP */ |
502 | close(fd); |
503 | } else { |
504 | /* "copy-on-write" or "anonymous" memory map */ |
505 | #ifdef MAP_ANONYMOUS |
506 | flags |= MAP_ANONYMOUS; |
507 | #else |
508 | if ((fd = open("/dev/zero" , O_RDWR | O_CLOEXEC)) < 0) { |
509 | GDKsyserror("MT_mremap: open(/dev/zero) failed\n" ); |
510 | fprintf(stderr, "= %s:%d: MT_mremap(%s,%p,%zu,%zu): open('/dev/zero') failed\n" , __FILE__, __LINE__, path?path:"NULL" , old_address, old_size, *new_size); |
511 | return NULL; |
512 | } |
513 | #endif |
514 | /* try to map an anonymous area as extent to the |
515 | * current map */ |
516 | p = mmap((char *) old_address + old_size, *new_size - old_size, |
517 | prot, flags, fd, 0); |
518 | /* no point trying a full map if this didn't work: |
519 | * there isn't enough space */ |
520 | if (p != MAP_FAILED) { |
521 | if (p == (char *) old_address + old_size) { |
522 | /* we got the requested address, make |
523 | * sure we return the correct (old) |
524 | * address */ |
525 | VALGRIND_RESIZEINPLACE_BLOCK(old_address, old_size, *new_size, 0); |
526 | p = old_address; |
527 | } else { |
528 | /* we got some other address: discard |
529 | * it and make full mmap */ |
530 | munmap(p, *new_size - old_size); |
531 | #ifdef HAVE_MREMAP |
532 | /* first get an area large enough for |
533 | * *new_size */ |
534 | p = mmap(NULL, *new_size, prot, flags, fd, 0); |
535 | if (p != MAP_FAILED) { |
536 | /* then overlay old mmap over new */ |
537 | void *q; |
538 | |
539 | q = mremap(old_address, old_size, |
540 | old_size, |
541 | MREMAP_FIXED | MREMAP_MAYMOVE, |
542 | p); |
543 | assert(q == p || q == MAP_FAILED); |
544 | if (q == MAP_FAILED) { |
545 | /* we didn't expect this... */ |
546 | munmap(p, *new_size); |
547 | p = MAP_FAILED; |
548 | } |
549 | #ifdef HAVE_VALGRIND |
550 | else { |
551 | VALGRIND_FREELIKE_BLOCK(old_size, 0); |
552 | VALGRIND_MALLOCLIKE_BLOCK(p, *new_size, 0, 1); |
553 | } |
554 | #endif |
555 | } |
556 | #else |
557 | p = MAP_FAILED; |
558 | if (path == NULL || |
559 | *new_size <= GDK_mmap_minsize_persistent) { |
560 | /* size not too big yet or |
561 | * anonymous, try to make new |
562 | * anonymous mmap and copy |
563 | * data over */ |
564 | p = mmap(NULL, *new_size, prot, flags, |
565 | fd, 0); |
566 | if (p != MAP_FAILED) { |
567 | VALGRIND_MALLOCLIKE_BLOCK(p, *new_size, 0, 0); |
568 | memcpy(p, old_address, |
569 | old_size); |
570 | munmap(old_address, old_size); |
571 | VALGRIND_FREELIKE_BLOCK(old_address, 0); |
572 | } |
573 | /* if it failed, try alternative */ |
574 | } |
575 | if (p == MAP_FAILED && path != NULL) { |
576 | #ifdef HAVE_POSIX_FALLOCATE |
577 | int rt; |
578 | #endif |
579 | /* write data to disk, then |
580 | * mmap it to new address */ |
581 | if (fd >= 0) |
582 | close(fd); |
583 | p = malloc(strlen(path) + 5); |
584 | if (p == NULL){ |
585 | GDKsyserror("MT_mremap: malloc() failed\n" ); |
586 | fprintf(stderr, "= %s:%d: MT_mremap(%s,%p,%zu,%zu): fd < 0\n" , __FILE__, __LINE__, path, old_address, old_size, *new_size); |
587 | return NULL; |
588 | } |
589 | |
590 | strcat(strcpy(p, path), ".tmp" ); |
591 | fd = open(p, O_RDWR | O_CREAT | O_CLOEXEC, |
592 | MONETDB_MODE); |
593 | if (fd < 0) { |
594 | GDKsyserror("MT_mremap: open(%s) failed\n" , (char *) p); |
595 | free(p); |
596 | fprintf(stderr, "= %s:%d: MT_mremap(%s,%p,%zu,%zu): fd < 0\n" , __FILE__, __LINE__, path, old_address, old_size, *new_size); |
597 | return NULL; |
598 | } |
599 | free(p); |
600 | if (write(fd, old_address, |
601 | old_size) < 0 || |
602 | #ifdef HAVE_FALLOCATE |
603 | /* prefer Linux-specific |
604 | * fallocate over standard |
605 | * posix_fallocate, since |
606 | * glibc uses a rather |
607 | * slow method of |
608 | * allocating the file if |
609 | * the file system doesn't |
610 | * support the operation, |
611 | * we just use ftruncate |
612 | * in that case */ |
613 | (fallocate(fd, 0, (off_t) old_size, (off_t) *new_size - (off_t) old_size) < 0 && (errno != EOPNOTSUPP || ftruncate(fd, (off_t) *new_size) < 0)) |
614 | #else |
615 | #ifdef HAVE_POSIX_FALLOCATE |
616 | /* posix_fallocate returns |
617 | * error number on |
618 | * failure, not -1, and if |
619 | * it returns EINVAL, the |
620 | * underlying file system |
621 | * may not support the |
622 | * operation, so we then |
623 | * need to try |
624 | * ftruncate */ |
625 | ((rt = posix_fallocate(fd, (off_t) old_size, (off_t) *new_size - (off_t) old_size)) == EINVAL ? ftruncate(fd, (off_t) *new_size) < 0 : rt != 0) |
626 | #else |
627 | ftruncate(fd, (off_t) *new_size) < 0 |
628 | #endif |
629 | #endif |
630 | ) { |
631 | int err = errno, other; |
632 | /* extending failed: |
633 | * free any disk space |
634 | * allocated in the |
635 | * process */ |
636 | other = ftruncate(fd, (off_t) old_size); |
637 | (void) other; /* silence compiler warning for ignoring result of ftruncate */ |
638 | errno = err; /* restore for error message */ |
639 | GDKsyserror("MT_mremap: growing file failed\n" ); |
640 | close(fd); |
641 | fprintf(stderr, |
642 | "= %s:%d: MT_mremap(%s,%p,%zu,%zu): write() or " |
643 | #ifdef HAVE_FALLOCATE |
644 | "fallocate()" |
645 | #else |
646 | #ifdef HAVE_POSIX_FALLOCATE |
647 | "posix_fallocate()" |
648 | #else |
649 | "ftruncate()" |
650 | #endif |
651 | #endif |
652 | " failed\n" , __FILE__, __LINE__, path, old_address, old_size, *new_size); |
653 | return NULL; |
654 | } |
655 | p = mmap(NULL, *new_size, prot, flags, |
656 | fd, 0); |
657 | if (p != MAP_FAILED) { |
658 | VALGRIND_MALLOCLIKE_BLOCK(p, *new_size, 0, 1); |
659 | munmap(old_address, old_size); |
660 | VALGRIND_FREELIKE_BLOCK(old_address, 0); |
661 | } |
662 | } |
663 | #endif /* HAVE_MREMAP */ |
664 | } |
665 | } |
666 | if (p == MAP_FAILED) |
667 | GDKsyserror("MT_mremap: mmap failed\n" ); |
668 | if (fd >= 0) |
669 | close(fd); |
670 | } |
671 | #ifdef MMAP_DEBUG |
672 | fprintf(stderr, "MT_mremap(%s,%p,%zu,%zu) -> %p%s\n" , path?path:"NULL" , old_address, old_size, *new_size, p, path && mode & MMAP_COPY ? " private" : "" ); |
673 | #endif |
674 | if (p == MAP_FAILED) |
675 | fprintf(stderr, "= %s:%d: MT_mremap(%s,%p,%zu,%zu): p == MAP_FAILED\n" , __FILE__, __LINE__, path?path:"NULL" , old_address, old_size, *new_size); |
676 | return p == MAP_FAILED ? NULL : p; |
677 | } |
678 | |
679 | int |
680 | MT_msync(void *p, size_t len) |
681 | { |
682 | int ret = msync(p, len, MS_SYNC); |
683 | |
684 | if (ret < 0) |
685 | GDKsyserror("MT_msync: msync failed\n" ); |
686 | #ifdef MMAP_DEBUG |
687 | fprintf(stderr, |
688 | "#msync(%p,%zu,MS_SYNC) = %d\n" , |
689 | p, len, ret); |
690 | #endif |
691 | return ret; |
692 | } |
693 | |
694 | bool |
695 | MT_path_absolute(const char *pathname) |
696 | { |
697 | return (*pathname == DIR_SEP); |
698 | } |
699 | |
700 | #ifdef HAVE_DLFCN_H |
701 | # include <dlfcn.h> |
702 | #endif |
703 | |
704 | void * |
705 | mdlopen(const char *library, int mode) |
706 | { |
707 | (void) library; |
708 | return dlopen(NULL, mode); |
709 | } |
710 | |
711 | #else /* WIN32 native */ |
712 | |
713 | #ifndef BUFSIZ |
714 | #define BUFSIZ 1024 |
715 | #endif |
716 | |
717 | #undef _errno |
718 | #undef stat |
719 | #undef rmdir |
720 | #undef mkdir |
721 | |
722 | #include <windows.h> |
723 | |
724 | #ifdef _MSC_VER |
725 | #include <io.h> |
726 | #endif /* _MSC_VER */ |
727 | #include <Psapi.h> |
728 | |
729 | #define MT_SMALLBLOCK 256 |
730 | |
731 | static LONG WINAPI |
732 | MT_ignore_exceptions(struct _EXCEPTION_POINTERS *ExceptionInfo) |
733 | { |
734 | (void) ExceptionInfo; |
735 | return EXCEPTION_EXECUTE_HANDLER; |
736 | } |
737 | |
738 | void |
739 | MT_init_posix(void) |
740 | { |
741 | SetUnhandledExceptionFilter(MT_ignore_exceptions); |
742 | } |
743 | |
744 | size_t |
745 | MT_getrss(void) |
746 | { |
747 | PROCESS_MEMORY_COUNTERS ctr; |
748 | if (GetProcessMemoryInfo(GetCurrentProcess(), &ctr, sizeof(ctr))) |
749 | return ctr.WorkingSetSize; |
750 | return 0; |
751 | } |
752 | |
753 | /* Windows mmap keeps a global list of base addresses for complex |
754 | * (remapped) memory maps the reason is that each remapped segment |
755 | * needs to be unmapped separately in the end. */ |
756 | |
757 | void * |
758 | MT_mmap(const char *path, int mode, size_t len) |
759 | { |
760 | DWORD mode0 = FILE_READ_ATTRIBUTES | FILE_READ_DATA; |
761 | DWORD mode1 = FILE_SHARE_READ | FILE_SHARE_WRITE; |
762 | DWORD mode2 = mode & MMAP_ADVISE; |
763 | DWORD mode3 = PAGE_READONLY; |
764 | int mode4 = FILE_MAP_READ; |
765 | SECURITY_ATTRIBUTES sa; |
766 | HANDLE h1, h2; |
767 | void *ret; |
768 | |
769 | if (mode & MMAP_WRITE) { |
770 | mode0 |= FILE_APPEND_DATA | FILE_WRITE_ATTRIBUTES | FILE_WRITE_DATA; |
771 | } |
772 | if (mode2 == MMAP_RANDOM || mode2 == MMAP_DONTNEED) { |
773 | mode2 = FILE_FLAG_RANDOM_ACCESS; |
774 | } else if (mode2 == MMAP_SEQUENTIAL || mode2 == MMAP_WILLNEED) { |
775 | mode2 = FILE_FLAG_SEQUENTIAL_SCAN; |
776 | } else { |
777 | mode2 = FILE_FLAG_NO_BUFFERING; |
778 | } |
779 | if (mode & MMAP_SYNC) { |
780 | mode2 |= FILE_FLAG_WRITE_THROUGH; |
781 | } |
782 | if (mode & MMAP_COPY) { |
783 | mode3 = PAGE_WRITECOPY; |
784 | mode4 = FILE_MAP_COPY; |
785 | } else if (mode & MMAP_WRITE) { |
786 | mode3 = PAGE_READWRITE; |
787 | mode4 = FILE_MAP_WRITE; |
788 | } |
789 | sa.nLength = sizeof(SECURITY_ATTRIBUTES); |
790 | sa.bInheritHandle = TRUE; |
791 | sa.lpSecurityDescriptor = 0; |
792 | |
793 | h1 = CreateFile(path, mode0, mode1, &sa, OPEN_ALWAYS, mode2, NULL); |
794 | if (h1 == INVALID_HANDLE_VALUE) { |
795 | (void) SetFileAttributes(path, FILE_ATTRIBUTE_NORMAL); |
796 | h1 = CreateFile(path, mode0, mode1, &sa, OPEN_ALWAYS, mode2, NULL); |
797 | if (h1 == INVALID_HANDLE_VALUE) { |
798 | GDKwinerror("MT_mmap: CreateFile('%s', %lu, %lu, &sa, %lu, %lu, NULL) failed\n" , |
799 | path, mode0, mode1, (DWORD) OPEN_ALWAYS, mode2); |
800 | return NULL; |
801 | } |
802 | } |
803 | |
804 | h2 = CreateFileMapping(h1, &sa, mode3, (DWORD) (((__int64) len >> 32) & LL_CONSTANT(0xFFFFFFFF)), (DWORD) (len & LL_CONSTANT(0xFFFFFFFF)), NULL); |
805 | if (h2 == NULL) { |
806 | GDKwinerror("MT_mmap: CreateFileMapping(%p, &sa, %lu, %lu, %lu, NULL) failed\n" , |
807 | h1, mode3, |
808 | (DWORD) (((__int64) len >> 32) & LL_CONSTANT(0xFFFFFFFF)), |
809 | (DWORD) (len & LL_CONSTANT(0xFFFFFFFF))); |
810 | CloseHandle(h1); |
811 | return NULL; |
812 | } |
813 | CloseHandle(h1); |
814 | |
815 | ret = MapViewOfFileEx(h2, mode4, (DWORD) 0, (DWORD) 0, len, NULL); |
816 | if (ret == NULL) |
817 | errno = winerror(GetLastError()); |
818 | CloseHandle(h2); |
819 | |
820 | return ret; |
821 | } |
822 | |
823 | int |
824 | MT_munmap(void *p, size_t dummy) |
825 | { |
826 | int ret; |
827 | |
828 | (void) dummy; |
829 | /* Windows' UnmapViewOfFile returns success!=0, error== 0, |
830 | * while Unix's munmap returns success==0, error==-1. */ |
831 | ret = UnmapViewOfFile(p); |
832 | if (ret == 0) { |
833 | GDKwinerror("MT_munmap failed\n" ); |
834 | return -1; |
835 | } |
836 | return 0; |
837 | } |
838 | |
839 | void * |
840 | MT_mremap(const char *path, int mode, void *old_address, size_t old_size, size_t *new_size) |
841 | { |
842 | void *p; |
843 | |
844 | /* doesn't make sense for us to extend read-only memory map */ |
845 | assert(mode & MMAP_WRITABLE); |
846 | |
847 | /* round up to multiple of page size */ |
848 | *new_size = (*new_size + GDK_mmap_pagesize - 1) & ~(GDK_mmap_pagesize - 1); |
849 | |
850 | if (old_size >= *new_size) { |
851 | *new_size = old_size; |
852 | return old_address; /* don't bother shrinking */ |
853 | } |
854 | if (GDKextend(path, *new_size) != GDK_SUCCEED) { |
855 | fprintf(stderr, "= %s:%d: MT_mremap(%s,%p,%zu,%zu): GDKextend() failed\n" , __FILE__, __LINE__, path?path:"NULL" , old_address, old_size, *new_size); |
856 | return NULL; |
857 | } |
858 | if (path && !(mode & MMAP_COPY)) |
859 | MT_munmap(old_address, old_size); |
860 | p = MT_mmap(path, mode, *new_size); |
861 | if (p != NULL && (path == NULL || (mode & MMAP_COPY))) { |
862 | memcpy(p, old_address, old_size); |
863 | MT_munmap(old_address, old_size); |
864 | } |
865 | #ifdef MMAP_DEBUG |
866 | fprintf(stderr, "MT_mremap(%s,%p,%zu,%zu) -> %p\n" , path?path:"NULL" , old_address, old_size, *new_size, p); |
867 | #endif |
868 | if (p == NULL) |
869 | fprintf(stderr, "= %s:%d: MT_mremap(%s,%p,%zu,%zu): p == NULL\n" , __FILE__, __LINE__, path?path:"NULL" , old_address, old_size, *new_size); |
870 | return p; |
871 | } |
872 | |
873 | int |
874 | MT_msync(void *p, size_t len) |
875 | { |
876 | int ret; |
877 | |
878 | /* Windows' FlushViewOfFile returns success!=0, error== 0, |
879 | * while Unix's munmap returns success==0, error==-1. */ |
880 | ret = FlushViewOfFile(p, len); |
881 | if (ret == 0) { |
882 | GDKwinerror("MT_msync: FlushViewOfFile failed\n" ); |
883 | return -1; |
884 | } |
885 | return 0; |
886 | } |
887 | |
888 | bool |
889 | MT_path_absolute(const char *pathname) |
890 | { |
891 | /* drive letter, colon, directory separator */ |
892 | return (((('a' <= pathname[0] && pathname[0] <= 'z') || |
893 | ('A' <= pathname[0] && pathname[0] <= 'Z')) && |
894 | pathname[1] == ':' && |
895 | (pathname[2] == '/' || pathname[2] == '\\')) || |
896 | (pathname[0] == '\\' && pathname[1] == '\\')); |
897 | } |
898 | |
899 | #ifndef HAVE_GETTIMEOFDAY |
900 | static int nodays[12] = { 31, 28, 31, 30, 31, 30, 31, 31, 30, 31, 30, 31 }; |
901 | |
902 | #define LEAPYEAR(y) ((((y)%4)==0 && ((y)%100)!=0) || ((y)%400)==0) |
903 | #define NODAYS(m,y) (((m)!=2)?nodays[(m)-1]:LEAPYEAR(y)?29:28) |
904 | |
905 | int |
906 | gettimeofday(struct timeval *tv, int *ignore_zone) |
907 | { |
908 | unsigned int year, day, month; |
909 | SYSTEMTIME st; |
910 | |
911 | (void) ignore_zone; |
912 | GetSystemTime(&st); |
913 | day = 0; |
914 | for (year = 1970; year < st.wYear; year++) |
915 | day += LEAPYEAR(year) ? 366 : 365; |
916 | |
917 | for (month = 1; month < st.wMonth; month++) |
918 | day += NODAYS(month, st.wYear); |
919 | |
920 | day += st.wDay; |
921 | tv->tv_sec = 60 * (day * 24 * 60 + st.wMinute) + st.wSecond; |
922 | tv->tv_usec = 1000 * st.wMilliseconds; |
923 | return 0; |
924 | } |
925 | #endif |
926 | |
927 | void * |
928 | mdlopen(const char *file, int mode) |
929 | { |
930 | return dlopen(file, mode); |
931 | } |
932 | |
933 | void * |
934 | dlopen(const char *file, int mode) |
935 | { |
936 | (void) mode; |
937 | if (file != NULL) { |
938 | return (void *) LoadLibrary(file); |
939 | } |
940 | return GetModuleHandle(NULL); |
941 | } |
942 | |
943 | int |
944 | dlclose(void *handle) |
945 | { |
946 | if (handle != NULL) { |
947 | return FreeLibrary((HINSTANCE) handle); |
948 | } |
949 | return -1; |
950 | } |
951 | |
952 | void * |
953 | dlsym(void *handle, const char *name) |
954 | { |
955 | if (handle != NULL) { |
956 | return (void *) GetProcAddress((HINSTANCE) handle, name); |
957 | } |
958 | return NULL; |
959 | } |
960 | |
961 | char * |
962 | dlerror(void) |
963 | { |
964 | static char msg[1024]; |
965 | |
966 | FormatMessage(FORMAT_MESSAGE_FROM_SYSTEM, NULL, GetLastError(), 0, msg, sizeof(msg), NULL); |
967 | return msg; |
968 | } |
969 | |
970 | /* dir manipulations fail in WIN32 if file name contains trailing |
971 | * slashes; work around this */ |
972 | static char * |
973 | reduce_dir_name(const char *src, char *dst, size_t cap) |
974 | { |
975 | size_t len = strlen(src); |
976 | char *buf = dst; |
977 | |
978 | if (len >= cap) |
979 | buf = malloc(len + 1); |
980 | if (buf == NULL) |
981 | return NULL; |
982 | while (--len > 0 && src[len - 1] != ':' && src[len] == DIR_SEP) |
983 | ; |
984 | for (buf[++len] = 0; len > 0; buf[len] = src[len]) |
985 | len--; |
986 | return buf; |
987 | } |
988 | |
989 | #undef _stat64 |
990 | int |
991 | win_stat(const char *pathname, struct _stat64 *st) |
992 | { |
993 | char buf[128], *p = reduce_dir_name(pathname, buf, sizeof(buf)); |
994 | int ret; |
995 | |
996 | if (p == NULL) |
997 | return -1; |
998 | ret = _stat64(p, st); |
999 | if (p != buf) |
1000 | free(p); |
1001 | return ret; |
1002 | } |
1003 | |
1004 | int |
1005 | win_rmdir(const char *pathname) |
1006 | { |
1007 | char buf[128], *p = reduce_dir_name(pathname, buf, sizeof(buf)); |
1008 | int ret; |
1009 | |
1010 | if (p == NULL) |
1011 | return -1; |
1012 | ret = _rmdir(p); |
1013 | if (ret < 0 && errno != ENOENT) { |
1014 | /* it could be the <expletive deleted> indexing |
1015 | * service which prevents us from doing what we have a |
1016 | * right to do, so try again (once) */ |
1017 | IODEBUG fprintf(stderr, "retry rmdir %s\n" , pathname); |
1018 | MT_sleep_ms(100); /* wait a little */ |
1019 | ret = _rmdir(p); |
1020 | } |
1021 | if (p != buf) |
1022 | free(p); |
1023 | return ret; |
1024 | } |
1025 | |
1026 | int |
1027 | win_unlink(const char *pathname) |
1028 | { |
1029 | int ret = _unlink(pathname); |
1030 | if (ret < 0) { |
1031 | /* Vista is paranoid: we cannot delete read-only files |
1032 | * owned by ourselves. Vista somehow also sets these |
1033 | * files to read-only. |
1034 | */ |
1035 | (void) SetFileAttributes(pathname, FILE_ATTRIBUTE_NORMAL); |
1036 | ret = _unlink(pathname); |
1037 | } |
1038 | if (ret < 0 && errno != ENOENT) { |
1039 | /* it could be the <expletive deleted> indexing |
1040 | * service which prevents us from doing what we have a |
1041 | * right to do, so try again (once) */ |
1042 | IODEBUG fprintf(stderr, "retry unlink %s\n" , pathname); |
1043 | MT_sleep_ms(100); /* wait a little */ |
1044 | ret = _unlink(pathname); |
1045 | } |
1046 | return ret; |
1047 | } |
1048 | |
1049 | #undef rename |
1050 | int |
1051 | win_rename(const char *old, const char *dst) |
1052 | { |
1053 | int ret; |
1054 | |
1055 | ret = rename(old, dst); |
1056 | if (ret == 0 || (ret < 0 && errno == ENOENT)) |
1057 | return ret; |
1058 | if (ret < 0 && errno == EEXIST) { |
1059 | (void) win_unlink(dst); |
1060 | ret = rename(old, dst); |
1061 | } |
1062 | |
1063 | if (ret < 0 && errno != ENOENT) { |
1064 | /* it could be the <expletive deleted> indexing |
1065 | * service which prevents us from doing what we have a |
1066 | * right to do, so try again (once) */ |
1067 | IODEBUG fprintf(stderr, "#retry rename %s %s\n" , old, dst); |
1068 | MT_sleep_ms(100); /* wait a little */ |
1069 | ret = rename(old, dst); |
1070 | } |
1071 | return ret; |
1072 | } |
1073 | |
1074 | int |
1075 | win_mkdir(const char *pathname, const int mode) |
1076 | { |
1077 | char buf[128], *p = reduce_dir_name(pathname, buf, sizeof(buf)); |
1078 | int ret; |
1079 | |
1080 | (void) mode; |
1081 | if (p == NULL) |
1082 | return -1; |
1083 | ret = _mkdir(p); |
1084 | if (p != buf) |
1085 | free(p); |
1086 | return ret; |
1087 | } |
1088 | #endif |
1089 | |
1090 | void |
1091 | MT_sleep_ms(unsigned int ms) |
1092 | { |
1093 | #ifdef NATIVE_WIN32 |
1094 | Sleep(ms); |
1095 | #else |
1096 | #ifdef HAVE_NANOSLEEP |
1097 | (void) nanosleep(&(struct timespec) {.tv_sec = ms / 1000, |
1098 | .tv_nsec = ms == 1 ? 1000 : (long) (ms % 1000) * 1000000,}, |
1099 | NULL); |
1100 | #else |
1101 | (void) select(0, NULL, NULL, NULL, |
1102 | &(struct timeval) {.tv_sec = ms / 1000, |
1103 | .tv_usec = ms == 1 ? 1 : (ms % 1000) * 1000,}); |
1104 | #endif |
1105 | #endif |
1106 | } |
1107 | |