FD.io VPP  v18.04-17-g3a0d853
Vector Packet Processing
svm.c
Go to the documentation of this file.
1 /*
2  *------------------------------------------------------------------
3  * svm.c - shared VM allocation, mmap(...MAP_FIXED...)
4  * library
5  *
6  * Copyright (c) 2009 Cisco and/or its affiliates.
7  * Licensed under the Apache License, Version 2.0 (the "License");
8  * you may not use this file except in compliance with the License.
9  * You may obtain a copy of the License at:
10  *
11  * http://www.apache.org/licenses/LICENSE-2.0
12  *
13  * Unless required by applicable law or agreed to in writing, software
14  * distributed under the License is distributed on an "AS IS" BASIS,
15  * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
16  * See the License for the specific language governing permissions and
17  * limitations under the License.
18  *------------------------------------------------------------------
19  */
20 
21 #include <stdio.h>
22 #include <stdlib.h>
23 #include <sys/types.h>
24 #include <sys/mman.h>
25 #include <sys/stat.h>
26 #include <netinet/in.h>
27 #include <signal.h>
28 #include <pthread.h>
29 #include <unistd.h>
30 #include <time.h>
31 #include <fcntl.h>
32 #include <string.h>
33 #include <vppinfra/clib.h>
34 #include <vppinfra/vec.h>
35 #include <vppinfra/hash.h>
36 #include <vppinfra/bitmap.h>
37 #include <vppinfra/fifo.h>
38 #include <vppinfra/time.h>
39 #include <vppinfra/mheap.h>
40 #include <vppinfra/heap.h>
41 #include <vppinfra/pool.h>
42 #include <vppinfra/format.h>
43 
44 #include "svm.h"
45 
47 static int root_rp_refcount;
48 
49 #define MAXLOCK 2
50 static pthread_mutex_t *mutexes_held[MAXLOCK];
51 static int nheld;
52 
55 {
56  return root_rp;
57 }
58 
59 #define MUTEX_DEBUG
60 
61 u64
63 {
64 #if __aarch64__
65  /* On AArch64 VA space can have different size, from 36 to 48 bits.
66  Here we are trying to detect VA bits by parsing /proc/self/maps
67  address ranges */
68  int fd;
69  unformat_input_t input;
70  u64 start, end = 0;
71  u8 bits = 0;
72 
73  if ((fd = open ("/proc/self/maps", 0)) < 0)
74  clib_unix_error ("open '/proc/self/maps'");
75 
76  unformat_init_clib_file (&input, fd);
78  {
79  if (unformat (&input, "%llx-%llx", &start, &end))
80  end--;
81  unformat_skip_line (&input);
82  }
83  unformat_free (&input);
84  close (fd);
85 
86  count_leading_zeros (bits, end);
87  bits = 64 - bits;
88  if (bits >= 36 && bits <= 48)
89  return ((1ul << bits) / 4) - (2 * SVM_GLOBAL_REGION_SIZE);
90  else
91  clib_unix_error ("unexpected va bits '%u'", bits);
92 #endif
93 
94  /* default value */
95  return 0x30000000;
96 }
97 
98 static void
99 region_lock (svm_region_t * rp, int tag)
100 {
101  pthread_mutex_lock (&rp->mutex);
102 #ifdef MUTEX_DEBUG
103  rp->mutex_owner_pid = getpid ();
104  rp->mutex_owner_tag = tag;
105 #endif
106  ASSERT (nheld < MAXLOCK);
107  /*
108  * Keep score of held mutexes so we can try to exit
109  * cleanly if the world comes to an end at the worst possible
110  * moment
111  */
112  mutexes_held[nheld++] = &rp->mutex;
113 }
114 
115 static void
117 {
118  int i, j;
119 #ifdef MUTEX_DEBUG
120  rp->mutex_owner_pid = 0;
121  rp->mutex_owner_tag = 0;
122 #endif
123 
124  for (i = nheld - 1; i >= 0; i--)
125  {
126  if (mutexes_held[i] == &rp->mutex)
127  {
128  for (j = i; j < MAXLOCK - 1; j++)
129  mutexes_held[j] = mutexes_held[j + 1];
130  nheld--;
131  goto found;
132  }
133  }
134  ASSERT (0);
135 
136 found:
138  pthread_mutex_unlock (&rp->mutex);
139 }
140 
141 
142 static u8 *
143 format_svm_flags (u8 * s, va_list * args)
144 {
145  uword f = va_arg (*args, uword);
146 
147  if (f & SVM_FLAGS_MHEAP)
148  s = format (s, "MHEAP ");
149  if (f & SVM_FLAGS_FILE)
150  s = format (s, "FILE ");
151  if (f & SVM_FLAGS_NODATA)
152  s = format (s, "NODATA ");
153  if (f & SVM_FLAGS_NEED_DATA_INIT)
154  s = format (s, "INIT ");
155 
156  return (s);
157 }
158 
159 static u8 *
160 format_svm_size (u8 * s, va_list * args)
161 {
162  uword size = va_arg (*args, uword);
163 
164  if (size >= (1 << 20))
165  {
166  s = format (s, "(%d mb)", size >> 20);
167  }
168  else if (size >= (1 << 10))
169  {
170  s = format (s, "(%d kb)", size >> 10);
171  }
172  else
173  {
174  s = format (s, "(%d bytes)", size);
175  }
176  return (s);
177 }
178 
179 u8 *
180 format_svm_region (u8 * s, va_list * args)
181 {
182  svm_region_t *rp = va_arg (*args, svm_region_t *);
183  int verbose = va_arg (*args, int);
184  int i;
185  uword lo, hi;
186 
187  s = format (s, "%s: base va 0x%x size 0x%x %U\n",
188  rp->region_name, rp->virtual_base,
190  s = format (s, " user_ctx 0x%x, bitmap_size %d\n",
191  rp->user_ctx, rp->bitmap_size);
192 
193  if (verbose)
194  {
195  s = format (s, " flags: 0x%x %U\n", rp->flags,
196  format_svm_flags, rp->flags);
197  s = format (s,
198  " region_heap 0x%x data_base 0x%x data_heap 0x%x\n",
199  rp->region_heap, rp->data_base, rp->data_heap);
200  }
201 
202  s = format (s, " %d clients, pids: ", vec_len (rp->client_pids));
203 
204  for (i = 0; i < vec_len (rp->client_pids); i++)
205  s = format (s, "%d ", rp->client_pids[i]);
206 
207  s = format (s, "\n");
208 
209  if (verbose)
210  {
211  lo = hi = ~0;
212 
213  s = format (s, " VM in use: ");
214 
215  for (i = 0; i < rp->bitmap_size; i++)
216  {
217  if (clib_bitmap_get_no_check (rp->bitmap, i) != 0)
218  {
219  if (lo == ~0)
220  {
221  hi = lo = rp->virtual_base + i * MMAP_PAGESIZE;
222  }
223  else
224  {
225  hi = rp->virtual_base + i * MMAP_PAGESIZE;
226  }
227  }
228  else
229  {
230  if (lo != ~0)
231  {
232  hi = rp->virtual_base + i * MMAP_PAGESIZE - 1;
233  s = format (s, " 0x%x - 0x%x (%dk)\n", lo, hi,
234  (hi - lo) >> 10);
235  lo = hi = ~0;
236  }
237  }
238  }
239  s = format (s, " rgn heap stats: %U", format_mheap,
240  rp->region_heap, 0);
241  if ((rp->flags & SVM_FLAGS_MHEAP) && rp->data_heap)
242  {
243  s = format (s, "\n data heap stats: %U", format_mheap,
244  rp->data_heap, 1);
245  }
246  s = format (s, "\n");
247  }
248 
249  return (s);
250 }
251 
252 /*
253  * rnd_pagesize
254  * Round to a pagesize multiple, presumably 4k works
255  */
256 static u64
258 {
259  u64 rv;
260 
261  rv = (size + (MMAP_PAGESIZE - 1)) & ~(MMAP_PAGESIZE - 1);
262  return (rv);
263 }
264 
265 /*
266  * svm_data_region_setup
267  */
268 static int
270 {
271  int fd;
272  u8 junk = 0;
273  uword map_size;
274 
275  map_size = rp->virtual_size - (MMAP_PAGESIZE +
276  (a->pvt_heap_size ? a->pvt_heap_size :
278 
279  if (a->flags & SVM_FLAGS_FILE)
280  {
281  struct stat statb;
282 
283  fd = open (a->backing_file, O_RDWR | O_CREAT, 0777);
284 
285  if (fd < 0)
286  {
287  clib_unix_warning ("open");
288  return -1;
289  }
290 
291  if (fstat (fd, &statb) < 0)
292  {
293  clib_unix_warning ("fstat");
294  close (fd);
295  return -2;
296  }
297 
298  if (statb.st_mode & S_IFREG)
299  {
300  if (statb.st_size == 0)
301  {
302  if (lseek (fd, map_size, SEEK_SET) == (off_t) - 1)
303  {
304  clib_unix_warning ("seek region size");
305  close (fd);
306  return -3;
307  }
308  if (write (fd, &junk, 1) != 1)
309  {
310  clib_unix_warning ("set region size");
311  close (fd);
312  return -3;
313  }
314  }
315  else
316  {
317  map_size = rnd_pagesize (statb.st_size);
318  }
319  }
320  else
321  {
322  map_size = a->backing_mmap_size;
323  }
324 
325  ASSERT (map_size <= rp->virtual_size -
327 
328  if (mmap (rp->data_base, map_size, PROT_READ | PROT_WRITE,
329  MAP_SHARED | MAP_FIXED, fd, 0) == MAP_FAILED)
330  {
331  clib_unix_warning ("mmap");
332  close (fd);
333  return -3;
334  }
335  close (fd);
336  rp->backing_file = (char *) format (0, "%s\0", a->backing_file);
337  rp->flags |= SVM_FLAGS_FILE;
338  }
339 
340  if (a->flags & SVM_FLAGS_MHEAP)
341  {
342  rp->data_heap =
343  mheap_alloc_with_flags ((void *) (rp->data_base), map_size,
345  rp->flags |= SVM_FLAGS_MHEAP;
346  }
347  return 0;
348 }
349 
350 static int
352 {
353  int fd;
354  u8 junk = 0;
355  uword map_size;
356  struct stat statb;
357 
358  map_size = rp->virtual_size -
361 
362  if (a->flags & SVM_FLAGS_FILE)
363  {
364 
365  fd = open (a->backing_file, O_RDWR, 0777);
366 
367  if (fd < 0)
368  {
369  clib_unix_warning ("open");
370  return -1;
371  }
372 
373  if (fstat (fd, &statb) < 0)
374  {
375  clib_unix_warning ("fstat");
376  close (fd);
377  return -2;
378  }
379 
380  if (statb.st_mode & S_IFREG)
381  {
382  if (statb.st_size == 0)
383  {
384  if (lseek (fd, map_size, SEEK_SET) == (off_t) - 1)
385  {
386  clib_unix_warning ("seek region size");
387  close (fd);
388  return -3;
389  }
390  if (write (fd, &junk, 1) != 1)
391  {
392  clib_unix_warning ("set region size");
393  close (fd);
394  return -3;
395  }
396  }
397  else
398  {
399  map_size = rnd_pagesize (statb.st_size);
400  }
401  }
402  else
403  {
404  map_size = a->backing_mmap_size;
405  }
406 
407  ASSERT (map_size <= rp->virtual_size
408  - (MMAP_PAGESIZE
409  +
411 
412  if (mmap (rp->data_base, map_size, PROT_READ | PROT_WRITE,
413  MAP_SHARED | MAP_FIXED, fd, 0) == MAP_FAILED)
414  {
415  clib_unix_warning ("mmap");
416  close (fd);
417  return -3;
418  }
419  close (fd);
420  }
421  return 0;
422 }
423 
424 u8 *
426 {
427  u8 *path;
428  u8 *shm_name;
429  u8 *split_point;
430  u8 *mkdir_arg = 0;
431  int root_path_offset = 0;
432  int name_offset = 0;
433 
434  if (a->root_path)
435  {
436  /* Tolerate present or absent slashes */
437  if (a->root_path[0] == '/')
438  root_path_offset++;
439 
440  /* create the root_path under /dev/shm
441  iterate through path creating directories */
442 
443  path = format (0, "/dev/shm/%s%c", &a->root_path[root_path_offset], 0);
444  split_point = path + 1;
445  vec_add1 (mkdir_arg, '-');
446 
447  while (*split_point)
448  {
449  while (*split_point && *split_point != '/')
450  {
451  vec_add1 (mkdir_arg, *split_point);
452  split_point++;
453  }
454  vec_add1 (mkdir_arg, 0);
455 
456  /* ready to descend another level */
457  mkdir_arg[vec_len (mkdir_arg) - 1] = '-';
458  split_point++;
459  }
460  vec_free (mkdir_arg);
461  vec_free (path);
462 
463  if (a->name[0] == '/')
464  name_offset = 1;
465 
466  shm_name = format (0, "/%s-%s%c", &a->root_path[root_path_offset],
467  &a->name[name_offset], 0);
468  }
469  else
470  shm_name = format (0, "%s%c", a->name, 0);
471  return (shm_name);
472 }
473 
474 void
476 {
477  pthread_mutexattr_t attr;
478  pthread_condattr_t cattr;
479  int nbits, words, bit;
480  int overhead_space;
481  void *oldheap;
482  uword data_base;
483  ASSERT (rp);
484  int rv;
485 
486  memset (rp, 0, sizeof (*rp));
487 
488  if (pthread_mutexattr_init (&attr))
489  clib_unix_warning ("mutexattr_init");
490 
491  if (pthread_mutexattr_setpshared (&attr, PTHREAD_PROCESS_SHARED))
492  clib_unix_warning ("mutexattr_setpshared");
493 
494  if (pthread_mutex_init (&rp->mutex, &attr))
495  clib_unix_warning ("mutex_init");
496 
497  if (pthread_mutexattr_destroy (&attr))
498  clib_unix_warning ("mutexattr_destroy");
499 
500  if (pthread_condattr_init (&cattr))
501  clib_unix_warning ("condattr_init");
502 
503  if (pthread_condattr_setpshared (&cattr, PTHREAD_PROCESS_SHARED))
504  clib_unix_warning ("condattr_setpshared");
505 
506  if (pthread_cond_init (&rp->condvar, &cattr))
507  clib_unix_warning ("cond_init");
508 
509  if (pthread_condattr_destroy (&cattr))
510  clib_unix_warning ("condattr_destroy");
511 
512  region_lock (rp, 1);
513 
514  rp->virtual_base = a->baseva;
515  rp->virtual_size = a->size;
516 
517  rp->region_heap =
519  (a->baseva + MMAP_PAGESIZE, void *),
520  (a->pvt_heap_size !=
523  oldheap = svm_push_pvt_heap (rp);
524 
525  rp->region_name = (char *) format (0, "%s%c", a->name, 0);
526  vec_add1 (rp->client_pids, getpid ());
527 
528  nbits = rp->virtual_size / MMAP_PAGESIZE;
529 
530  ASSERT (nbits > 0);
531  rp->bitmap_size = nbits;
532  words = (nbits + BITS (uword) - 1) / BITS (uword);
533  vec_validate (rp->bitmap, words - 1);
534 
535  overhead_space = MMAP_PAGESIZE /* header */ +
537 
538  bit = 0;
539  data_base = (uword) rp->virtual_base;
540 
541  if (a->flags & SVM_FLAGS_NODATA)
543 
544  do
545  {
546  clib_bitmap_set_no_check (rp->bitmap, bit, 1);
547  bit++;
548  overhead_space -= MMAP_PAGESIZE;
549  data_base += MMAP_PAGESIZE;
550  }
551  while (overhead_space > 0);
552 
553  rp->data_base = (void *) data_base;
554 
555  /*
556  * Note: although the POSIX spec guarantees that only one
557  * process enters this block, we have to play games
558  * to hold off clients until e.g. the mutex is ready
559  */
560  rp->version = SVM_VERSION;
561 
562  /* setup the data portion of the region */
563 
564  rv = svm_data_region_create (a, rp);
565  if (rv)
566  {
567  clib_warning ("data_region_create: %d", rv);
568  }
569 
570  region_unlock (rp);
571 
572  svm_pop_heap (oldheap);
573 }
574 
575 /*
576  * svm_map_region
577  */
578 void *
580 {
581  int svm_fd;
582  svm_region_t *rp;
583  int deadman = 0;
584  u8 junk = 0;
585  void *oldheap;
586  int rv;
587  int pid_holding_region_lock;
588  u8 *shm_name;
589  int dead_region_recovery = 0;
590  int time_left;
591  struct stat stat;
592  struct timespec ts, tsrem;
593 
594  ASSERT ((a->size & ~(MMAP_PAGESIZE - 1)) == a->size);
595  ASSERT (a->name);
596 
597  shm_name = shm_name_from_svm_map_region_args (a);
598 
599  if (CLIB_DEBUG > 1)
600  clib_warning ("[%d] map region %s: shm_open (%s)",
601  getpid (), a->name, shm_name);
602 
603  svm_fd = shm_open ((char *) shm_name, O_RDWR | O_CREAT | O_EXCL, 0777);
604 
605  if (svm_fd >= 0)
606  {
607  if (fchmod (svm_fd, S_IRUSR | S_IWUSR | S_IRGRP | S_IWGRP) < 0)
608  clib_unix_warning ("segment chmod");
609  /* This turns out to fail harmlessly if the client starts first */
610  if (fchown (svm_fd, a->uid, a->gid) < 0)
611  clib_unix_warning ("segment chown [ok if client starts first]");
612 
613  vec_free (shm_name);
614 
615  if (lseek (svm_fd, a->size, SEEK_SET) == (off_t) - 1)
616  {
617  clib_warning ("seek region size");
618  close (svm_fd);
619  return (0);
620  }
621  if (write (svm_fd, &junk, 1) != 1)
622  {
623  clib_warning ("set region size");
624  close (svm_fd);
625  return (0);
626  }
627 
628  rp = mmap (uword_to_pointer (a->baseva, void *), a->size,
629  PROT_READ | PROT_WRITE, MAP_SHARED | MAP_FIXED, svm_fd, 0);
630 
631  if (rp == (svm_region_t *) MAP_FAILED)
632  {
633  clib_unix_warning ("mmap create");
634  close (svm_fd);
635  return (0);
636  }
637  close (svm_fd);
638 
640 
641  return ((void *) rp);
642  }
643  else
644  {
645  svm_fd = shm_open ((char *) shm_name, O_RDWR, 0777);
646 
647  vec_free (shm_name);
648 
649  if (svm_fd < 0)
650  {
651  perror ("svm_region_map(mmap open)");
652  return (0);
653  }
654 
655  time_left = 20;
656  while (1)
657  {
658  if (0 != fstat (svm_fd, &stat))
659  {
660  clib_warning ("fstat failed: %d", errno);
661  close (svm_fd);
662  return (0);
663  }
664  if (stat.st_size > 0)
665  {
666  break;
667  }
668  if (0 == time_left)
669  {
670  clib_warning ("waiting for resize of shm file timed out");
671  close (svm_fd);
672  return (0);
673  }
674  ts.tv_sec = 0;
675  ts.tv_nsec = 100000000;
676  while (nanosleep (&ts, &tsrem) < 0)
677  ts = tsrem;
678  time_left--;
679  }
680 
681  rp = mmap (0, MMAP_PAGESIZE,
682  PROT_READ | PROT_WRITE, MAP_SHARED, svm_fd, 0);
683 
684  if (rp == (svm_region_t *) MAP_FAILED)
685  {
686  close (svm_fd);
687  clib_warning ("mmap");
688  return (0);
689  }
690  /*
691  * We lost the footrace to create this region; make sure
692  * the winner has crossed the finish line.
693  */
694  while (rp->version == 0 && deadman++ < 5)
695  {
696  sleep (1);
697  }
698 
699  /*
700  * <bleep>-ed?
701  */
702  if (rp->version == 0)
703  {
704  clib_warning ("rp->version %d not %d", rp->version, SVM_VERSION);
705  close (svm_fd);
706  munmap (rp, a->size);
707  return (0);
708  }
709  /* Remap now that the region has been placed */
710  a->baseva = rp->virtual_base;
711  a->size = rp->virtual_size;
712  munmap (rp, MMAP_PAGESIZE);
713 
714  rp = (void *) mmap (uword_to_pointer (a->baseva, void *), a->size,
715  PROT_READ | PROT_WRITE,
716  MAP_SHARED | MAP_FIXED, svm_fd, 0);
717  if ((uword) rp == (uword) MAP_FAILED)
718  {
719  clib_unix_warning ("mmap");
720  close (svm_fd);
721  return (0);
722  }
723 
724  close (svm_fd);
725 
726  if ((uword) rp != rp->virtual_base)
727  {
728  clib_warning ("mmap botch");
729  }
730 
731  /*
732  * Try to fix the region mutex if it is held by
733  * a dead process
734  */
735  pid_holding_region_lock = rp->mutex_owner_pid;
736  if (pid_holding_region_lock && kill (pid_holding_region_lock, 0) < 0)
737  {
739  ("region %s mutex held by dead pid %d, tag %d, force unlock",
740  rp->region_name, pid_holding_region_lock, rp->mutex_owner_tag);
741  /* owner pid is nonexistent */
742  rp->mutex.__data.__owner = 0;
743  rp->mutex.__data.__lock = 0;
744  dead_region_recovery = 1;
745  }
746 
747  if (dead_region_recovery)
748  clib_warning ("recovery: attempt to re-lock region");
749 
750  region_lock (rp, 2);
751  oldheap = svm_push_pvt_heap (rp);
752  vec_add1 (rp->client_pids, getpid ());
753 
754  if (dead_region_recovery)
755  clib_warning ("recovery: attempt svm_data_region_map");
756 
757  rv = svm_data_region_map (a, rp);
758  if (rv)
759  {
760  clib_warning ("data_region_map: %d", rv);
761  }
762 
763  if (dead_region_recovery)
764  clib_warning ("unlock and continue");
765 
766  region_unlock (rp);
767 
768  svm_pop_heap (oldheap);
769 
770  return ((void *) rp);
771 
772  }
773  return 0; /* NOTREACHED */
774 }
775 
776 static void
778 {
779  int i;
780  for (i = 0; i < nheld; i++)
781  {
782  pthread_mutex_unlock (mutexes_held[i]);
783  }
784 }
785 
786 static int
788 {
789  svm_region_t *rp;
790  u64 ticks = clib_cpu_time_now ();
791  uword randomize_baseva;
792 
793  /* guard against klutz calls */
794  if (root_rp)
795  return -1;
796 
798 
799  atexit (svm_mutex_cleanup);
800 
801  /* Randomize the shared-VM base at init time */
802  if (MMAP_PAGESIZE <= (4 << 10))
803  randomize_baseva = (ticks & 15) * MMAP_PAGESIZE;
804  else
805  randomize_baseva = (ticks & 3) * MMAP_PAGESIZE;
806 
807  a->baseva += randomize_baseva;
808 
809  rp = svm_map_region (a);
810  if (!rp)
811  return -1;
812 
813  region_lock (rp, 3);
814 
815  /* Set up the main region data structures */
817  {
818  svm_main_region_t *mp = 0;
819  void *oldheap;
820 
822 
823  oldheap = svm_push_pvt_heap (rp);
824  vec_validate (mp, 0);
825  mp->name_hash = hash_create_string (0, sizeof (uword));
826  mp->root_path = a->root_path ? format (0, "%s%c", a->root_path, 0) : 0;
827  mp->uid = a->uid;
828  mp->gid = a->gid;
829  rp->data_base = mp;
830  svm_pop_heap (oldheap);
831  }
832  region_unlock (rp);
833  root_rp = rp;
834 
835  return 0;
836 }
837 
838 void
840 {
841  svm_map_region_args_t _a, *a = &_a;
842 
843  memset (a, 0, sizeof (*a));
844  a->root_path = 0;
848  a->flags = SVM_FLAGS_NODATA;
849  a->uid = 0;
850  a->gid = 0;
851 
853 }
854 
855 int
856 svm_region_init_chroot (const char *root_path)
857 {
858  svm_map_region_args_t _a, *a = &_a;
859 
860  memset (a, 0, sizeof (*a));
861  a->root_path = root_path;
865  a->flags = SVM_FLAGS_NODATA;
866  a->uid = 0;
867  a->gid = 0;
868 
869  return svm_region_init_internal (a);
870 }
871 
872 void
873 svm_region_init_chroot_uid_gid (const char *root_path, int uid, int gid)
874 {
875  svm_map_region_args_t _a, *a = &_a;
876 
877  memset (a, 0, sizeof (*a));
878  a->root_path = root_path;
882  a->flags = SVM_FLAGS_NODATA;
883  a->uid = uid;
884  a->gid = gid;
885 
887 }
888 
889 void
891 {
893 }
894 
895 void *
897 {
898  svm_main_region_t *mp;
899  svm_region_t *rp;
900  uword need_nbits;
901  int index, i;
902  void *oldheap;
903  uword *p;
904  u8 *name;
905  svm_subregion_t *subp;
906 
907  ASSERT (root_rp);
908 
909  a->size += MMAP_PAGESIZE +
911  a->size = rnd_pagesize (a->size);
912 
913  region_lock (root_rp, 4);
914  oldheap = svm_push_pvt_heap (root_rp);
915  mp = root_rp->data_base;
916 
917  ASSERT (mp);
918 
919  /* Map the named region from the correct chroot environment */
920  if (a->root_path == NULL)
921  a->root_path = (char *) mp->root_path;
922 
923  /*
924  * See if this region is already known. If it is, we're
925  * almost done...
926  */
927  p = hash_get_mem (mp->name_hash, a->name);
928 
929  if (p)
930  {
931  rp = svm_map_region (a);
932  region_unlock (root_rp);
933  svm_pop_heap (oldheap);
934  return rp;
935  }
936 
937  /* Create the region. */
938  ASSERT ((a->size & ~(MMAP_PAGESIZE - 1)) == a->size);
939 
940  need_nbits = a->size / MMAP_PAGESIZE;
941 
942  index = 1; /* $$$ fixme, figure out how many bit to really skip */
943 
944  /*
945  * Scan the virtual space allocation bitmap, looking for a large
946  * enough chunk
947  */
948  do
949  {
950  if (clib_bitmap_get_no_check (root_rp->bitmap, index) == 0)
951  {
952  for (i = 0; i < (need_nbits - 1); i++)
953  {
954  if (clib_bitmap_get_no_check (root_rp->bitmap, index + i) == 1)
955  {
956  index = index + i;
957  goto next;
958  }
959  }
960  break;
961  }
962  index++;
963  next:;
964  }
965  while (index < root_rp->bitmap_size);
966 
967  /* Completely out of VM? */
968  if (index >= root_rp->bitmap_size)
969  {
970  clib_warning ("region %s: not enough VM to allocate 0x%llx (%lld)",
971  root_rp->region_name, a->size, a->size);
972  svm_pop_heap (oldheap);
973  region_unlock (root_rp);
974  return 0;
975  }
976 
977  /*
978  * Mark virtual space allocated
979  */
980 #if CLIB_DEBUG > 1
981  clib_warning ("set %d bits at index %d", need_nbits, index);
982 #endif
983 
984  for (i = 0; i < need_nbits; i++)
985  {
986  clib_bitmap_set_no_check (root_rp->bitmap, index + i, 1);
987  }
988 
989  /* Place this region where it goes... */
990  a->baseva = root_rp->virtual_base + index * MMAP_PAGESIZE;
991 
992  rp = svm_map_region (a);
993 
994  pool_get (mp->subregions, subp);
995  name = format (0, "%s%c", a->name, 0);
996  subp->subregion_name = name;
997 
998  hash_set_mem (mp->name_hash, name, subp - mp->subregions);
999 
1000  svm_pop_heap (oldheap);
1001 
1002  region_unlock (root_rp);
1003 
1004  return (rp);
1005 }
1006 
1007 void
1009 {
1010  svm_map_region_args_t _a, *a = &_a;
1011  svm_main_region_t *mp;
1012  u8 *shm_name;
1013 
1014  ASSERT (root_rp);
1015  ASSERT (rp);
1017 
1018  mp = root_rp->data_base;
1019  ASSERT (mp);
1020 
1021  a->root_path = (char *) mp->root_path;
1022  a->name = rp->region_name;
1023  shm_name = shm_name_from_svm_map_region_args (a);
1024  if (CLIB_DEBUG > 1)
1025  clib_warning ("[%d] shm_unlink (%s)", getpid (), shm_name);
1026  shm_unlink ((const char *) shm_name);
1027  vec_free (shm_name);
1028 }
1029 
1030 /*
1031  * svm_region_unmap
1032  *
1033  * Let go of the indicated region. If the calling process
1034  * is the last customer, throw it away completely.
1035  * The root region mutex guarantees atomicity with respect to
1036  * a new region client showing up at the wrong moment.
1037  */
1038 void
1039 svm_region_unmap_internal (void *rp_arg, u8 is_client)
1040 {
1041  int i, mypid = getpid ();
1042  int nclients_left;
1043  void *oldheap;
1044  uword virtual_base, virtual_size;
1045  svm_region_t *rp = rp_arg;
1046  char *name;
1047 
1048  /*
1049  * If we take a signal while holding one or more shared-memory
1050  * mutexes, we may end up back here from an otherwise
1051  * benign exit handler. Bail out to avoid a recursive
1052  * mutex screw-up.
1053  */
1054  if (nheld)
1055  return;
1056 
1057  ASSERT (rp);
1058  ASSERT (root_rp);
1059 
1060  if (CLIB_DEBUG > 1)
1061  clib_warning ("[%d] unmap region %s", getpid (), rp->region_name);
1062 
1063  region_lock (root_rp, 5);
1064  region_lock (rp, 6);
1065 
1066  oldheap = svm_push_pvt_heap (rp); /* nb vec_delete() in the loop */
1067 
1068  /* Remove the caller from the list of mappers */
1069  for (i = 0; i < vec_len (rp->client_pids); i++)
1070  {
1071  if (rp->client_pids[i] == mypid)
1072  {
1073  vec_delete (rp->client_pids, 1, i);
1074  goto found;
1075  }
1076  }
1077  clib_warning ("pid %d AWOL", mypid);
1078 
1079 found:
1080 
1081  svm_pop_heap (oldheap);
1082 
1083  nclients_left = vec_len (rp->client_pids);
1084  virtual_base = rp->virtual_base;
1085  virtual_size = rp->virtual_size;
1086 
1087  if (nclients_left == 0)
1088  {
1089  int index, nbits, i;
1090  svm_main_region_t *mp;
1091  uword *p;
1092  svm_subregion_t *subp;
1093 
1094  /* Kill the region, last guy on his way out */
1095 
1096  oldheap = svm_push_pvt_heap (root_rp);
1097  name = vec_dup (rp->region_name);
1098 
1099  virtual_base = rp->virtual_base;
1100  virtual_size = rp->virtual_size;
1101 
1102  /* Figure out which bits to clear in the root region bitmap */
1103  index = (virtual_base - root_rp->virtual_base) / MMAP_PAGESIZE;
1104 
1105  nbits = (virtual_size + MMAP_PAGESIZE - 1) / MMAP_PAGESIZE;
1106 
1107 #if CLIB_DEBUG > 1
1108  clib_warning ("clear %d bits at index %d", nbits, index);
1109 #endif
1110  /* Give back the allocated VM */
1111  for (i = 0; i < nbits; i++)
1112  {
1113  clib_bitmap_set_no_check (root_rp->bitmap, index + i, 0);
1114  }
1115 
1116  mp = root_rp->data_base;
1117 
1118  p = hash_get_mem (mp->name_hash, name);
1119 
1120  /* Better never happen ... */
1121  if (p == NULL)
1122  {
1123  region_unlock (rp);
1124  region_unlock (root_rp);
1125  svm_pop_heap (oldheap);
1126  clib_warning ("Region name '%s' not found?", name);
1127  return;
1128  }
1129 
1130  /* Remove from the root region subregion pool */
1131  subp = mp->subregions + p[0];
1132  pool_put (mp->subregions, subp);
1133 
1134  hash_unset_mem (mp->name_hash, name);
1135 
1136  vec_free (name);
1137 
1138  region_unlock (rp);
1139 
1140  /* If a client asks for the cleanup, don't unlink the backing
1141  * file since we can't tell if it has been recreated. */
1142  if (!is_client)
1143  svm_region_unlink (rp);
1144 
1145  munmap ((void *) virtual_base, virtual_size);
1146  region_unlock (root_rp);
1147  svm_pop_heap (oldheap);
1148  return;
1149  }
1150 
1151  region_unlock (rp);
1152  region_unlock (root_rp);
1153 
1154  munmap ((void *) virtual_base, virtual_size);
1155 }
1156 
1157 void
1158 svm_region_unmap (void *rp_arg)
1159 {
1160  svm_region_unmap_internal (rp_arg, 0 /* is_client */ );
1161 }
1162 
1163 void
1165 {
1166  svm_region_unmap_internal (rp_arg, 1 /* is_client */ );
1167 }
1168 
1169 /*
1170  * svm_region_exit
1171  */
1172 static void
1174 {
1175  void *oldheap;
1176  int i, mypid = getpid ();
1177  uword virtual_base, virtual_size;
1178 
1179  /* It felt so nice we did it twice... */
1180  if (root_rp == 0)
1181  return;
1182 
1183  if (--root_rp_refcount > 0)
1184  return;
1185 
1186  /*
1187  * If we take a signal while holding one or more shared-memory
1188  * mutexes, we may end up back here from an otherwise
1189  * benign exit handler. Bail out to avoid a recursive
1190  * mutex screw-up.
1191  */
1192  if (nheld)
1193  return;
1194 
1195  region_lock (root_rp, 7);
1196  oldheap = svm_push_pvt_heap (root_rp);
1197 
1198  virtual_base = root_rp->virtual_base;
1199  virtual_size = root_rp->virtual_size;
1200 
1201  for (i = 0; i < vec_len (root_rp->client_pids); i++)
1202  {
1203  if (root_rp->client_pids[i] == mypid)
1204  {
1205  vec_delete (root_rp->client_pids, 1, i);
1206  goto found;
1207  }
1208  }
1209  clib_warning ("pid %d AWOL", mypid);
1210 
1211 found:
1212 
1213  if (!is_client && vec_len (root_rp->client_pids) == 0)
1214  svm_region_unlink (root_rp);
1215 
1216  region_unlock (root_rp);
1217  svm_pop_heap (oldheap);
1218 
1219  root_rp = 0;
1220  munmap ((void *) virtual_base, virtual_size);
1221 }
1222 
1223 void
1225 {
1226  svm_region_exit_internal (0 /* is_client */ );
1227 }
1228 
1229 void
1231 {
1232  svm_region_exit_internal (1 /* is_client */ );
1233 }
1234 
1235 void
1237 {
1238  int j;
1239  int mypid = getpid ();
1240  void *oldheap;
1241 
1242  for (j = 0; j < vec_len (rp->client_pids); j++)
1243  {
1244  if (mypid == rp->client_pids[j])
1245  continue;
1246  if (rp->client_pids[j] && (kill (rp->client_pids[j], 0) < 0))
1247  {
1248  clib_warning ("%s: cleanup ghost pid %d",
1249  rp->region_name, rp->client_pids[j]);
1250  /* nb: client vec in rp->region_heap */
1251  oldheap = svm_push_pvt_heap (rp);
1252  vec_delete (rp->client_pids, 1, j);
1253  j--;
1254  svm_pop_heap (oldheap);
1255  }
1256  }
1257 }
1258 
1259 
1260 /*
1261  * Scan svm regions for dead clients
1262  */
1263 void
1264 svm_client_scan (const char *root_path)
1265 {
1266  int i, j;
1267  svm_main_region_t *mp;
1268  svm_map_region_args_t *a = 0;
1270  svm_region_t *rp;
1271  svm_subregion_t *subp;
1272  u8 *name = 0;
1273  u8 **svm_names = 0;
1274  void *oldheap;
1275  int mypid = getpid ();
1276 
1277  vec_validate (a, 0);
1278 
1279  svm_region_init_chroot (root_path);
1280 
1281  root_rp = svm_get_root_rp ();
1282 
1283  pthread_mutex_lock (&root_rp->mutex);
1284 
1285  mp = root_rp->data_base;
1286 
1287  for (j = 0; j < vec_len (root_rp->client_pids); j++)
1288  {
1289  if (mypid == root_rp->client_pids[j])
1290  continue;
1291  if (root_rp->client_pids[j] && (kill (root_rp->client_pids[j], 0) < 0))
1292  {
1293  clib_warning ("%s: cleanup ghost pid %d",
1294  root_rp->region_name, root_rp->client_pids[j]);
1295  /* nb: client vec in root_rp->region_heap */
1296  oldheap = svm_push_pvt_heap (root_rp);
1297  vec_delete (root_rp->client_pids, 1, j);
1298  j--;
1299  svm_pop_heap (oldheap);
1300  }
1301  }
1302 
1303  /*
1304  * Snapshoot names, can't hold root rp mutex across
1305  * find_or_create.
1306  */
1307  /* *INDENT-OFF* */
1308  pool_foreach (subp, mp->subregions, ({
1309  name = vec_dup (subp->subregion_name);
1310  vec_add1(svm_names, name);
1311  }));
1312  /* *INDENT-ON* */
1313 
1314  pthread_mutex_unlock (&root_rp->mutex);
1315 
1316  for (i = 0; i < vec_len (svm_names); i++)
1317  {
1318  vec_validate (a, 0);
1319  a->root_path = root_path;
1320  a->name = (char *) svm_names[i];
1321  rp = svm_region_find_or_create (a);
1322  if (rp)
1323  {
1324  pthread_mutex_lock (&rp->mutex);
1325 
1327 
1328  pthread_mutex_unlock (&rp->mutex);
1329  svm_region_unmap (rp);
1330  vec_free (svm_names[i]);
1331  }
1332  vec_free (a);
1333  }
1334  vec_free (svm_names);
1335 
1336  svm_region_exit ();
1337 
1338  vec_free (a);
1339 }
1340 
1341 /*
1342  * fd.io coding-style-patch-verification: ON
1343  *
1344  * Local Variables:
1345  * eval: (c-set-style "gnu")
1346  * End:
1347  */
#define vec_validate(V, I)
Make sure vector is long enough for given index (no header, unspecified alignment) ...
Definition: vec.h:434
vmrglw vmrglh hi
void svm_region_init_chroot_uid_gid(const char *root_path, int uid, int gid)
Definition: svm.c:873
svm_region_t * svm_get_root_rp(void)
Definition: svm.c:54
#define SVM_GLOBAL_REGION_NAME
Definition: svm_common.h:87
const char * root_path
Definition: svm_common.h:67
static int nheld
Definition: svm.c:51
static void svm_pop_heap(void *oldheap)
Definition: svm.h:94
#define vec_c_string_is_terminated(V)
Test whether a vector is a NULL terminated c-string.
Definition: vec.h:984
a
Definition: bitmap.h:516
#define SVM_FLAGS_NODATA
Definition: svm_common.h:29
#define SVM_FLAGS_NEED_DATA_INIT
Definition: svm_common.h:30
void * svm_map_region(svm_map_region_args_t *a)
Definition: svm.c:579
Fixed length block allocator.
#define NULL
Definition: clib.h:55
#define vec_add1(V, E)
Add 1 element to end of vector (unspecified alignment).
Definition: vec.h:520
static u64 clib_cpu_time_now(void)
Definition: time.h:73
void svm_client_scan(const char *root_path)
Definition: svm.c:1264
uword virtual_base
Definition: svm_common.h:42
#define SVM_PVT_MHEAP_SIZE
Definition: svm_common.h:32
int i
void svm_region_unmap_client(void *rp_arg)
Definition: svm.c:1164
#define hash_set_mem(h, key, value)
Definition: hash.h:274
void svm_region_exit_client(void)
Definition: svm.c:1230
u8 * format(u8 *s, const char *fmt,...)
Definition: format.c:419
#define pool_get(P, E)
Allocate an object E from a pool P (unspecified alignment).
Definition: pool.h:227
static u8 * format_svm_size(u8 *s, va_list *args)
Definition: svm.c:160
static uword clib_bitmap_get_no_check(uword *ai, uword i)
Gets the ith bit value from a bitmap Does not sanity-check the bit position.
Definition: bitmap.h:212
u8 * format_mheap(u8 *s, va_list *va)
Definition: mheap.c:1162
void unformat_init_clib_file(unformat_input_t *input, int file_descriptor)
Definition: unformat.c:1058
#define clib_unix_error(format, args...)
Definition: error.h:65
static uword clib_bitmap_set_no_check(uword *a, uword i, uword new_value)
Sets the ith bit of a bitmap to new_value.
Definition: bitmap.h:141
#define SVM_VERSION
Definition: svm_common.h:25
void svm_region_exit(void)
Definition: svm.c:1224
static void unformat_skip_line(unformat_input_t *i)
Definition: format.h:220
void svm_region_init(void)
Definition: svm.c:839
#define pool_foreach(VAR, POOL, BODY)
Iterate through pool.
Definition: pool.h:440
uword * client_pids
Definition: svm_common.h:54
#define MHEAP_FLAG_DISABLE_VM
void * svm_region_find_or_create(svm_map_region_args_t *a)
Definition: svm.c:896
volatile void * user_ctx
Definition: svm_common.h:47
static u64 rnd_pagesize(u64 size)
Definition: svm.c:257
unsigned long u64
Definition: types.h:89
pthread_cond_t condvar
Definition: svm_common.h:38
u8 * format_svm_region(u8 *s, va_list *args)
Definition: svm.c:180
#define hash_create_string(elts, value_bytes)
Definition: hash.h:675
#define SVM_FLAGS_MHEAP
Definition: svm_common.h:27
static int svm_region_init_internal(svm_map_region_args_t *a)
Definition: svm.c:787
void * data_base
Definition: svm_common.h:45
u32 size
#define hash_unset_mem(h, key)
Definition: hash.h:290
lo
#define MAXLOCK
Definition: svm.c:49
struct _unformat_input_t unformat_input_t
#define pool_put(P, E)
Free an object E in pool P.
Definition: pool.h:273
#define vec_dup(V)
Return copy of vector (no header, no alignment)
Definition: vec.h:370
svm_subregion_t * subregions
Definition: svm_common.h:105
char * backing_file
Definition: svm_common.h:52
uword virtual_size
Definition: svm_common.h:43
void svm_region_init_args(svm_map_region_args_t *a)
Definition: svm.c:890
#define SVM_GLOBAL_REGION_SIZE
Definition: svm_common.h:86
char * region_name
Definition: svm_common.h:51
void svm_region_init_mapped_region(svm_map_region_args_t *a, svm_region_t *rp)
Definition: svm.c:475
#define uword_to_pointer(u, type)
Definition: types.h:136
static void * svm_push_pvt_heap(svm_region_t *rp)
Definition: svm.h:78
static int root_rp_refcount
Definition: svm.c:47
#define UNFORMAT_END_OF_INPUT
Definition: format.h:143
#define vec_free(V)
Free vector&#39;s memory (no header).
Definition: vec.h:336
#define clib_warning(format, args...)
Definition: error.h:59
static pthread_mutex_t * mutexes_held[MAXLOCK]
Definition: svm.c:50
u8 * shm_name_from_svm_map_region_args(svm_map_region_args_t *a)
Definition: svm.c:425
uword bitmap_size
Definition: svm_common.h:49
static int svm_data_region_map(svm_map_region_args_t *a, svm_region_t *rp)
Definition: svm.c:351
void * mheap_alloc_with_flags(void *memory, uword memory_size, uword flags)
Definition: mheap.c:869
#define ASSERT(truth)
#define vec_delete(V, N, M)
Delete N elements starting at element M.
Definition: vec.h:783
static void region_lock(svm_region_t *rp, int tag)
Definition: svm.c:99
volatile uword version
Definition: svm_common.h:36
u64 svm_get_global_region_base_va()
Definition: svm.c:62
Bitmaps built as vectors of machine words.
static void region_unlock(svm_region_t *rp)
Definition: svm.c:116
int mutex_owner_tag
Definition: svm_common.h:40
void svm_region_unmap_internal(void *rp_arg, u8 is_client)
Definition: svm.c:1039
#define MMAP_PAGESIZE
Definition: ssvm.h:42
u64 uword
Definition: types.h:112
void svm_region_unlink(svm_region_t *rp)
Definition: svm.c:1008
#define vec_len(v)
Number of elements in vector (rvalue-only, NULL tolerant)
unsigned char u8
Definition: types.h:56
const char * name
Definition: svm_common.h:68
static void unformat_free(unformat_input_t *i)
Definition: format.h:161
#define clib_unix_warning(format, args...)
Definition: error.h:68
void svm_region_unmap(void *rp_arg)
Definition: svm.c:1158
void svm_client_scan_this_region_nolock(svm_region_t *rp)
Definition: svm.c:1236
#define hash_get_mem(h, key)
Definition: hash.h:268
void * region_heap
Definition: svm_common.h:44
uword * bitmap
Definition: svm_common.h:50
#define CLIB_MEMORY_BARRIER()
Definition: clib.h:109
void * data_heap
Definition: svm_common.h:46
static int svm_data_region_create(svm_map_region_args_t *a, svm_region_t *rp)
Definition: svm.c:269
static u8 * format_svm_flags(u8 *s, va_list *args)
Definition: svm.c:143
int mutex_owner_pid
Definition: svm_common.h:39
#define BITS(x)
Definition: clib.h:58
uword flags
Definition: svm_common.h:41
static void svm_mutex_cleanup(void)
Definition: svm.c:777
uword unformat(unformat_input_t *i, const char *fmt,...)
Definition: unformat.c:972
pthread_mutex_t mutex
Definition: svm_common.h:37
#define SVM_FLAGS_FILE
Definition: svm_common.h:28
CLIB vectors are ubiquitous dynamically resized arrays with by user defined "headers".
static uword unformat_check_input(unformat_input_t *i)
Definition: format.h:169
static svm_region_t * root_rp
Definition: svm.c:46
int svm_region_init_chroot(const char *root_path)
Definition: svm.c:856
static void svm_region_exit_internal(u8 is_client)
Definition: svm.c:1173