FD.io VPP  v16.12-rc0-308-g931be3a
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 static void
62 region_lock (svm_region_t * rp, int tag)
63 {
64  pthread_mutex_lock (&rp->mutex);
65 #ifdef MUTEX_DEBUG
66  rp->mutex_owner_pid = getpid ();
67  rp->mutex_owner_tag = tag;
68 #endif
69  ASSERT (nheld < MAXLOCK);
70  /*
71  * Keep score of held mutexes so we can try to exit
72  * cleanly if the world comes to an end at the worst possible
73  * moment
74  */
75  mutexes_held[nheld++] = &rp->mutex;
76 }
77 
78 static void
80 {
81  int i, j;
82 #ifdef MUTEX_DEBUG
83  rp->mutex_owner_pid = 0;
84  rp->mutex_owner_tag = 0;
85 #endif
86 
87  for (i = nheld - 1; i >= 0; i--)
88  {
89  if (mutexes_held[i] == &rp->mutex)
90  {
91  for (j = i; j < MAXLOCK - 1; j++)
92  mutexes_held[j] = mutexes_held[j + 1];
93  nheld--;
94  goto found;
95  }
96  }
97  ASSERT (0);
98 
99 found:
101  pthread_mutex_unlock (&rp->mutex);
102 }
103 
104 
105 static u8 *
106 format_svm_flags (u8 * s, va_list * args)
107 {
108  uword f = va_arg (*args, uword);
109 
110  if (f & SVM_FLAGS_MHEAP)
111  s = format (s, "MHEAP ");
112  if (f & SVM_FLAGS_FILE)
113  s = format (s, "FILE ");
114  if (f & SVM_FLAGS_NODATA)
115  s = format (s, "NODATA ");
116  if (f & SVM_FLAGS_NEED_DATA_INIT)
117  s = format (s, "INIT ");
118 
119  return (s);
120 }
121 
122 static u8 *
123 format_svm_size (u8 * s, va_list * args)
124 {
125  uword size = va_arg (*args, uword);
126 
127  if (size >= (1 << 20))
128  {
129  s = format (s, "(%d mb)", size >> 20);
130  }
131  else if (size >= (1 << 10))
132  {
133  s = format (s, "(%d kb)", size >> 10);
134  }
135  else
136  {
137  s = format (s, "(%d bytes)", size);
138  }
139  return (s);
140 }
141 
142 u8 *
143 format_svm_region (u8 * s, va_list * args)
144 {
145  svm_region_t *rp = va_arg (*args, svm_region_t *);
146  int verbose = va_arg (*args, int);
147  int i;
148  uword lo, hi;
149 
150  s = format (s, "%s: base va 0x%x size 0x%x %U\n",
151  rp->region_name, rp->virtual_base,
153  s = format (s, " user_ctx 0x%x, bitmap_size %d\n",
154  rp->user_ctx, rp->bitmap_size);
155 
156  if (verbose)
157  {
158  s = format (s, " flags: 0x%x %U\n", rp->flags,
159  format_svm_flags, rp->flags);
160  s = format (s,
161  " region_heap 0x%x data_base 0x%x data_heap 0x%x\n",
162  rp->region_heap, rp->data_base, rp->data_heap);
163  }
164 
165  s = format (s, " %d clients, pids: ", vec_len (rp->client_pids));
166 
167  for (i = 0; i < vec_len (rp->client_pids); i++)
168  s = format (s, "%d ", rp->client_pids[i]);
169 
170  s = format (s, "\n");
171 
172  if (verbose)
173  {
174  lo = hi = ~0;
175 
176  s = format (s, " VM in use: ");
177 
178  for (i = 0; i < rp->bitmap_size; i++)
179  {
180  if (clib_bitmap_get_no_check (rp->bitmap, i) != 0)
181  {
182  if (lo == ~0)
183  {
184  hi = lo = rp->virtual_base + i * MMAP_PAGESIZE;
185  }
186  else
187  {
188  hi = rp->virtual_base + i * MMAP_PAGESIZE;
189  }
190  }
191  else
192  {
193  if (lo != ~0)
194  {
195  hi = rp->virtual_base + i * MMAP_PAGESIZE - 1;
196  s = format (s, " 0x%x - 0x%x (%dk)\n", lo, hi,
197  (hi - lo) >> 10);
198  lo = hi = ~0;
199  }
200  }
201  }
202  s = format (s, " rgn heap stats: %U", format_mheap,
203  rp->region_heap, 0);
204  if ((rp->flags & SVM_FLAGS_MHEAP) && rp->data_heap)
205  {
206  s = format (s, "\n data heap stats: %U", format_mheap,
207  rp->data_heap, 1);
208  }
209  s = format (s, "\n");
210  }
211 
212  return (s);
213 }
214 
215 /*
216  * rnd_pagesize
217  * Round to a pagesize multiple, presumably 4k works
218  */
219 static u64
221 {
222  u64 rv;
223 
224  rv = (size + (MMAP_PAGESIZE - 1)) & ~(MMAP_PAGESIZE - 1);
225  return (rv);
226 }
227 
228 /*
229  * svm_data_region_setup
230  */
231 static int
233 {
234  int fd;
235  u8 junk = 0;
236  uword map_size;
237 
238  map_size = rp->virtual_size - (MMAP_PAGESIZE +
239  (a->pvt_heap_size ? a->pvt_heap_size :
241 
242  if (a->flags & SVM_FLAGS_FILE)
243  {
244  struct stat statb;
245 
246  fd = open (a->backing_file, O_RDWR | O_CREAT, 0777);
247 
248  if (fd < 0)
249  {
250  clib_unix_warning ("open");
251  return -1;
252  }
253 
254  if (fstat (fd, &statb) < 0)
255  {
256  clib_unix_warning ("fstat");
257  close (fd);
258  return -2;
259  }
260 
261  if (statb.st_mode & S_IFREG)
262  {
263  if (statb.st_size == 0)
264  {
265  if (lseek (fd, map_size, SEEK_SET) == (off_t) - 1)
266  {
267  clib_unix_warning ("seek region size");
268  close (fd);
269  return -3;
270  }
271  if (write (fd, &junk, 1) != 1)
272  {
273  clib_unix_warning ("set region size");
274  close (fd);
275  return -3;
276  }
277  }
278  else
279  {
280  map_size = rnd_pagesize (statb.st_size);
281  }
282  }
283  else
284  {
285  map_size = a->backing_mmap_size;
286  }
287 
288  ASSERT (map_size <= rp->virtual_size -
290 
291  if (mmap (rp->data_base, map_size, PROT_READ | PROT_WRITE,
292  MAP_SHARED | MAP_FIXED, fd, 0) == MAP_FAILED)
293  {
294  clib_unix_warning ("mmap");
295  close (fd);
296  return -3;
297  }
298  close (fd);
299  rp->backing_file = (char *) format (0, "%s\0", a->backing_file);
300  rp->flags |= SVM_FLAGS_FILE;
301  }
302 
303  if (a->flags & SVM_FLAGS_MHEAP)
304  {
305  rp->data_heap =
306  mheap_alloc_with_flags ((void *) (rp->data_base), map_size,
308  rp->flags |= SVM_FLAGS_MHEAP;
309  }
310  return 0;
311 }
312 
313 static int
315 {
316  int fd;
317  u8 junk = 0;
318  uword map_size;
319  struct stat statb;
320 
321  map_size = rp->virtual_size -
324 
325  if (a->flags & SVM_FLAGS_FILE)
326  {
327 
328  fd = open (a->backing_file, O_RDWR, 0777);
329 
330  if (fd < 0)
331  {
332  clib_unix_warning ("open");
333  return -1;
334  }
335 
336  if (fstat (fd, &statb) < 0)
337  {
338  clib_unix_warning ("fstat");
339  close (fd);
340  return -2;
341  }
342 
343  if (statb.st_mode & S_IFREG)
344  {
345  if (statb.st_size == 0)
346  {
347  if (lseek (fd, map_size, SEEK_SET) == (off_t) - 1)
348  {
349  clib_unix_warning ("seek region size");
350  close (fd);
351  return -3;
352  }
353  if (write (fd, &junk, 1) != 1)
354  {
355  clib_unix_warning ("set region size");
356  close (fd);
357  return -3;
358  }
359  }
360  else
361  {
362  map_size = rnd_pagesize (statb.st_size);
363  }
364  }
365  else
366  {
367  map_size = a->backing_mmap_size;
368  }
369 
370  ASSERT (map_size <= rp->virtual_size
371  - (MMAP_PAGESIZE
372  +
374 
375  if (mmap (rp->data_base, map_size, PROT_READ | PROT_WRITE,
376  MAP_SHARED | MAP_FIXED, fd, 0) == MAP_FAILED)
377  {
378  clib_unix_warning ("mmap");
379  close (fd);
380  return -3;
381  }
382  close (fd);
383  }
384  return 0;
385 }
386 
387 u8 *
389 {
390  u8 *path;
391  u8 *shm_name;
392  u8 *split_point;
393  u8 *mkdir_arg = 0;
394  int root_path_offset = 0;
395  int name_offset = 0;
396 
397  if (a->root_path)
398  {
399  /* Tolerate present or absent slashes */
400  if (a->root_path[0] == '/')
401  root_path_offset++;
402 
403  /* create the root_path under /dev/shm
404  iterate through path creating directories */
405 
406  path = format (0, "/dev/shm/%s%c", &a->root_path[root_path_offset], 0);
407  split_point = path + 1;
408  vec_add1 (mkdir_arg, '-');
409 
410  while (*split_point)
411  {
412  while (*split_point && *split_point != '/')
413  {
414  vec_add1 (mkdir_arg, *split_point);
415  split_point++;
416  }
417  vec_add1 (mkdir_arg, 0);
418 
419  /* ready to descend another level */
420  mkdir_arg[vec_len (mkdir_arg) - 1] = '-';
421  split_point++;
422  }
423  vec_free (mkdir_arg);
424  vec_free (path);
425 
426  if (a->name[0] == '/')
427  name_offset = 1;
428 
429  shm_name = format (0, "/%s-%s%c", a->root_path,
430  &a->name[name_offset], 0);
431  }
432  else
433  shm_name = format (0, "%s%c", a->name, 0);
434  return (shm_name);
435 }
436 
437 /*
438  * svm_map_region
439  */
440 void *
442 {
443  int svm_fd;
444  svm_region_t *rp;
445  pthread_mutexattr_t attr;
446  pthread_condattr_t cattr;
447  int deadman = 0;
448  u8 junk = 0;
449  void *oldheap;
450  int overhead_space;
451  int rv;
452  uword data_base;
453  int nbits, words, bit;
454  int pid_holding_region_lock;
455  u8 *shm_name;
456  int dead_region_recovery = 0;
457  int time_left;
458  struct stat stat;
459  struct timespec ts, tsrem;
460 
461  if (CLIB_DEBUG > 1)
462  clib_warning ("[%d] map region %s", getpid (), a->name);
463 
464  ASSERT ((a->size & ~(MMAP_PAGESIZE - 1)) == a->size);
465  ASSERT (a->name);
466 
467  shm_name = shm_name_from_svm_map_region_args (a);
468 
469  svm_fd = shm_open ((char *) shm_name, O_RDWR | O_CREAT | O_EXCL, 0777);
470 
471  if (svm_fd >= 0)
472  {
473  if (fchmod (svm_fd, 0770) < 0)
474  clib_unix_warning ("segment chmod");
475  /* This turns out to fail harmlessly if the client starts first */
476  if (fchown (svm_fd, a->uid, a->gid) < 0)
477  clib_unix_warning ("segment chown [ok if client starts first]");
478 
479  vec_free (shm_name);
480 
481  if (lseek (svm_fd, a->size, SEEK_SET) == (off_t) - 1)
482  {
483  clib_warning ("seek region size");
484  close (svm_fd);
485  return (0);
486  }
487  if (write (svm_fd, &junk, 1) != 1)
488  {
489  clib_warning ("set region size");
490  close (svm_fd);
491  return (0);
492  }
493 
494  rp = mmap ((void *) a->baseva, a->size,
495  PROT_READ | PROT_WRITE, MAP_SHARED | MAP_FIXED, svm_fd, 0);
496 
497  if (rp == (svm_region_t *) MAP_FAILED)
498  {
499  clib_unix_warning ("mmap create");
500  close (svm_fd);
501  return (0);
502  }
503  close (svm_fd);
504  memset (rp, 0, sizeof (*rp));
505 
506  if (pthread_mutexattr_init (&attr))
507  clib_unix_warning ("mutexattr_init");
508 
509  if (pthread_mutexattr_setpshared (&attr, PTHREAD_PROCESS_SHARED))
510  clib_unix_warning ("mutexattr_setpshared");
511 
512  if (pthread_mutex_init (&rp->mutex, &attr))
513  clib_unix_warning ("mutex_init");
514 
515  if (pthread_mutexattr_destroy (&attr))
516  clib_unix_warning ("mutexattr_destroy");
517 
518  if (pthread_condattr_init (&cattr))
519  clib_unix_warning ("condattr_init");
520 
521  if (pthread_condattr_setpshared (&cattr, PTHREAD_PROCESS_SHARED))
522  clib_unix_warning ("condattr_setpshared");
523 
524  if (pthread_cond_init (&rp->condvar, &cattr))
525  clib_unix_warning ("cond_init");
526 
527  if (pthread_condattr_destroy (&cattr))
528  clib_unix_warning ("condattr_destroy");
529 
530  region_lock (rp, 1);
531 
532  rp->virtual_base = a->baseva;
533  rp->virtual_size = a->size;
534 
535  rp->region_heap =
537  (a->pvt_heap_size != 0) ?
540  oldheap = svm_push_pvt_heap (rp);
541 
542  rp->region_name = (char *) format (0, "%s%c", a->name, 0);
543  vec_add1 (rp->client_pids, getpid ());
544 
545  nbits = rp->virtual_size / MMAP_PAGESIZE;
546 
547  ASSERT (nbits > 0);
548  rp->bitmap_size = nbits;
549  words = (nbits + BITS (uword) - 1) / BITS (uword);
550  vec_validate (rp->bitmap, words - 1);
551 
552  overhead_space = MMAP_PAGESIZE /* header */ +
554 
555  bit = 0;
556  data_base = (uword) rp->virtual_base;
557 
558  if (a->flags & SVM_FLAGS_NODATA)
560 
561  do
562  {
563  clib_bitmap_set_no_check (rp->bitmap, bit, 1);
564  bit++;
565  overhead_space -= MMAP_PAGESIZE;
566  data_base += MMAP_PAGESIZE;
567  }
568  while (overhead_space > 0);
569 
570  rp->data_base = (void *) data_base;
571 
572  /*
573  * Note: although the POSIX spec guarantees that only one
574  * process enters this block, we have to play games
575  * to hold off clients until e.g. the mutex is ready
576  */
577  rp->version = SVM_VERSION;
578 
579  /* setup the data portion of the region */
580 
581  rv = svm_data_region_create (a, rp);
582  if (rv)
583  {
584  clib_warning ("data_region_create: %d", rv);
585  }
586 
587  region_unlock (rp);
588 
589  svm_pop_heap (oldheap);
590 
591  return ((void *) rp);
592  }
593  else
594  {
595  svm_fd = shm_open ((char *) shm_name, O_RDWR, 0777);
596 
597  vec_free (shm_name);
598 
599  if (svm_fd < 0)
600  {
601  perror ("svm_region_map(mmap open)");
602  return (0);
603  }
604 
605  time_left = 20;
606  while (1)
607  {
608  if (0 != fstat (svm_fd, &stat))
609  {
610  clib_warning ("fstat failed: %d", errno);
611  close (svm_fd);
612  return (0);
613  }
614  if (stat.st_size > 0)
615  {
616  break;
617  }
618  if (0 == time_left)
619  {
620  clib_warning ("waiting for resize of shm file timed out");
621  close (svm_fd);
622  return (0);
623  }
624  ts.tv_sec = 0;
625  ts.tv_nsec = 100000000;
626  while (nanosleep (&ts, &tsrem) < 0)
627  ts = tsrem;
628  time_left--;
629  }
630 
631  rp = mmap (0, MMAP_PAGESIZE,
632  PROT_READ | PROT_WRITE, MAP_SHARED, svm_fd, 0);
633 
634  if (rp == (svm_region_t *) MAP_FAILED)
635  {
636  close (svm_fd);
637  clib_warning ("mmap");
638  return (0);
639  }
640  /*
641  * We lost the footrace to create this region; make sure
642  * the winner has crossed the finish line.
643  */
644  while (rp->version == 0 && deadman++ < 5)
645  {
646  sleep (1);
647  }
648 
649  /*
650  * <bleep>-ed?
651  */
652  if (rp->version == 0)
653  {
654  clib_warning ("rp->version %d not %d", rp->version, SVM_VERSION);
655  close (svm_fd);
656  munmap (rp, a->size);
657  return (0);
658  }
659  /* Remap now that the region has been placed */
660  a->baseva = rp->virtual_base;
661  a->size = rp->virtual_size;
662  munmap (rp, MMAP_PAGESIZE);
663 
664  rp = (void *) mmap ((void *) a->baseva, a->size,
665  PROT_READ | PROT_WRITE,
666  MAP_SHARED | MAP_FIXED, svm_fd, 0);
667  if ((uword) rp == (uword) MAP_FAILED)
668  {
669  clib_unix_warning ("mmap");
670  close (svm_fd);
671  return (0);
672  }
673 
674  if ((uword) rp != rp->virtual_base)
675  {
676  clib_warning ("mmap botch");
677  }
678 
679  /*
680  * Try to fix the region mutex if it is held by
681  * a dead process
682  */
683  pid_holding_region_lock = rp->mutex_owner_pid;
684  if (pid_holding_region_lock && kill (pid_holding_region_lock, 0) < 0)
685  {
687  ("region %s mutex held by dead pid %d, tag %d, force unlock",
688  rp->region_name, pid_holding_region_lock, rp->mutex_owner_tag);
689  /* owner pid is nonexistent */
690  rp->mutex.__data.__owner = 0;
691  rp->mutex.__data.__lock = 0;
692  dead_region_recovery = 1;
693  }
694 
695  if (dead_region_recovery)
696  clib_warning ("recovery: attempt to re-lock region");
697 
698  region_lock (rp, 2);
699  oldheap = svm_push_pvt_heap (rp);
700  vec_add1 (rp->client_pids, getpid ());
701 
702  if (dead_region_recovery)
703  clib_warning ("recovery: attempt svm_data_region_map");
704 
705  rv = svm_data_region_map (a, rp);
706  if (rv)
707  {
708  clib_warning ("data_region_map: %d", rv);
709  }
710 
711  if (dead_region_recovery)
712  clib_warning ("unlock and continue");
713 
714  region_unlock (rp);
715 
716  svm_pop_heap (oldheap);
717 
718  return ((void *) rp);
719 
720  }
721  return 0; /* NOTREACHED */
722 }
723 
724 static void
726 {
727  int i;
728  for (i = 0; i < nheld; i++)
729  {
730  pthread_mutex_unlock (mutexes_held[i]);
731  }
732 }
733 
734 static void
736 {
737  svm_region_t *rp;
738  u64 ticks = clib_cpu_time_now ();
739  uword randomize_baseva;
740 
741  /* guard against klutz calls */
742  if (root_rp)
743  return;
744 
746 
747  atexit (svm_mutex_cleanup);
748 
749  /* Randomize the shared-VM base at init time */
750  if (MMAP_PAGESIZE <= (4 << 10))
751  randomize_baseva = (ticks & 15) * MMAP_PAGESIZE;
752  else
753  randomize_baseva = (ticks & 3) * MMAP_PAGESIZE;
754 
755  a->baseva += randomize_baseva;
756 
757  rp = svm_map_region (a);
758  ASSERT (rp);
759 
760  region_lock (rp, 3);
761 
762  /* Set up the main region data structures */
764  {
765  svm_main_region_t *mp = 0;
766  void *oldheap;
767 
769 
770  oldheap = svm_push_pvt_heap (rp);
771  vec_validate (mp, 0);
772  mp->name_hash = hash_create_string (0, sizeof (uword));
773  mp->root_path = a->root_path ? format (0, "%s%c", a->root_path, 0) : 0;
774  rp->data_base = mp;
775  svm_pop_heap (oldheap);
776  }
777  region_unlock (rp);
778  root_rp = rp;
779 }
780 
781 void
783 {
784  svm_map_region_args_t _a, *a = &_a;
785 
786  memset (a, 0, sizeof (*a));
787  a->root_path = 0;
791  a->flags = SVM_FLAGS_NODATA;
792  a->uid = 0;
793  a->gid = 0;
794 
796 }
797 
798 void
799 svm_region_init_chroot (char *root_path)
800 {
801  svm_map_region_args_t _a, *a = &_a;
802 
803  memset (a, 0, sizeof (*a));
804  a->root_path = root_path;
808  a->flags = SVM_FLAGS_NODATA;
809  a->uid = 0;
810  a->gid = 0;
811 
813 }
814 
815 void
816 svm_region_init_chroot_uid_gid (char *root_path, int uid, int gid)
817 {
818  svm_map_region_args_t _a, *a = &_a;
819 
820  memset (a, 0, sizeof (*a));
821  a->root_path = root_path;
825  a->flags = SVM_FLAGS_NODATA;
826  a->uid = uid;
827  a->gid = gid;
828 
830 }
831 
832 void
834 {
836 }
837 
838 void *
840 {
841  svm_main_region_t *mp;
842  svm_region_t *rp;
843  uword need_nbits;
844  int index, i;
845  void *oldheap;
846  uword *p;
847  u8 *name;
848  svm_subregion_t *subp;
849 
850  ASSERT (root_rp);
851 
852  a->size += MMAP_PAGESIZE +
854  a->size = rnd_pagesize (a->size);
855 
856  region_lock (root_rp, 4);
857  oldheap = svm_push_pvt_heap (root_rp);
858  mp = root_rp->data_base;
859 
860  ASSERT (mp);
861 
862  /* Map the named region from the correct chroot environment */
863  a->root_path = (char *) mp->root_path;
864 
865  /*
866  * See if this region is already known. If it is, we're
867  * almost done...
868  */
869  p = hash_get_mem (mp->name_hash, a->name);
870 
871  if (p)
872  {
873  rp = svm_map_region (a);
874  region_unlock (root_rp);
875  svm_pop_heap (oldheap);
876  return rp;
877  }
878 
879  /* Create the region. */
880  ASSERT ((a->size & ~(MMAP_PAGESIZE - 1)) == a->size);
881 
882  need_nbits = a->size / MMAP_PAGESIZE;
883 
884  index = 1; /* $$$ fixme, figure out how many bit to really skip */
885 
886  /*
887  * Scan the virtual space allocation bitmap, looking for a large
888  * enough chunk
889  */
890  do
891  {
892  if (clib_bitmap_get_no_check (root_rp->bitmap, index) == 0)
893  {
894  for (i = 0; i < (need_nbits - 1); i++)
895  {
896  if (clib_bitmap_get_no_check (root_rp->bitmap, index + i) == 1)
897  {
898  index = index + i;
899  goto next;
900  }
901  }
902  break;
903  }
904  index++;
905  next:;
906  }
907  while (index < root_rp->bitmap_size);
908 
909  /* Completely out of VM? */
910  if (index >= root_rp->bitmap_size)
911  {
912  clib_warning ("region %s: not enough VM to allocate 0x%llx (%lld)",
913  root_rp->region_name, a->size, a->size);
914  svm_pop_heap (oldheap);
915  region_unlock (root_rp);
916  return 0;
917  }
918 
919  /*
920  * Mark virtual space allocated
921  */
922 #if CLIB_DEBUG > 1
923  clib_warning ("set %d bits at index %d", need_nbits, index);
924 #endif
925 
926  for (i = 0; i < need_nbits; i++)
927  {
928  clib_bitmap_set_no_check (root_rp->bitmap, index + i, 1);
929  }
930 
931  /* Place this region where it goes... */
932  a->baseva = root_rp->virtual_base + index * MMAP_PAGESIZE;
933 
934  rp = svm_map_region (a);
935 
936  pool_get (mp->subregions, subp);
937  name = format (0, "%s%c", a->name, 0);
938  subp->subregion_name = name;
939 
940  hash_set_mem (mp->name_hash, name, subp - mp->subregions);
941 
942  svm_pop_heap (oldheap);
943 
944  region_unlock (root_rp);
945 
946  return (rp);
947 }
948 
949 /*
950  * svm_region_unmap
951  *
952  * Let go of the indicated region. If the calling process
953  * is the last customer, throw it away completely.
954  * The root region mutex guarantees atomicity with respect to
955  * a new region client showing up at the wrong moment.
956  */
957 void
958 svm_region_unmap (void *rp_arg)
959 {
960  int i, mypid = getpid ();
961  int nclients_left;
962  void *oldheap;
963  uword virtual_base, virtual_size;
964  svm_region_t *rp = rp_arg;
965  char *name;
966 
967  /*
968  * If we take a signal while holding one or more shared-memory
969  * mutexes, we may end up back here from an otherwise
970  * benign exit handler. Bail out to avoid a recursive
971  * mutex screw-up.
972  */
973  if (nheld)
974  return;
975 
976  ASSERT (rp);
977  ASSERT (root_rp);
978 
979  if (CLIB_DEBUG > 1)
980  clib_warning ("[%d] unmap region %s", getpid (), rp->region_name);
981 
982  region_lock (root_rp, 5);
983  region_lock (rp, 6);
984 
985  oldheap = svm_push_pvt_heap (rp); /* nb vec_delete() in the loop */
986 
987  /* Remove the caller from the list of mappers */
988  for (i = 0; i < vec_len (rp->client_pids); i++)
989  {
990  if (rp->client_pids[i] == mypid)
991  {
992  vec_delete (rp->client_pids, 1, i);
993  goto found;
994  }
995  }
996  clib_warning ("pid %d AWOL", mypid);
997 
998 found:
999 
1000  svm_pop_heap (oldheap);
1001 
1002  nclients_left = vec_len (rp->client_pids);
1003  virtual_base = rp->virtual_base;
1004  virtual_size = rp->virtual_size;
1005 
1006  if (nclients_left == 0)
1007  {
1008  int index, nbits, i;
1009  svm_main_region_t *mp;
1010  uword *p;
1011  svm_subregion_t *subp;
1012 
1013  /* Kill the region, last guy on his way out */
1014 
1015  oldheap = svm_push_pvt_heap (root_rp);
1016  name = vec_dup (rp->region_name);
1017 
1018  virtual_base = rp->virtual_base;
1019  virtual_size = rp->virtual_size;
1020 
1021  /* Figure out which bits to clear in the root region bitmap */
1022  index = (virtual_base - root_rp->virtual_base) / MMAP_PAGESIZE;
1023 
1024  nbits = (virtual_size + MMAP_PAGESIZE - 1) / MMAP_PAGESIZE;
1025 
1026 #if CLIB_DEBUG > 1
1027  clib_warning ("clear %d bits at index %d", nbits, index);
1028 #endif
1029  /* Give back the allocated VM */
1030  for (i = 0; i < nbits; i++)
1031  {
1032  clib_bitmap_set_no_check (root_rp->bitmap, index + i, 0);
1033  }
1034 
1035  mp = root_rp->data_base;
1036 
1037  p = hash_get_mem (mp->name_hash, name);
1038 
1039  /* Better never happen ... */
1040  if (p == NULL)
1041  {
1042  region_unlock (rp);
1043  region_unlock (root_rp);
1044  svm_pop_heap (oldheap);
1045  clib_warning ("Region name '%s' not found?", name);
1046  return;
1047  }
1048 
1049  /* Remove from the root region subregion pool */
1050  subp = mp->subregions + p[0];
1051  pool_put (mp->subregions, subp);
1052 
1053  hash_unset_mem (mp->name_hash, name);
1054 
1055  vec_free (name);
1056 
1057  region_unlock (rp);
1058  shm_unlink (rp->region_name);
1059  munmap ((void *) virtual_base, virtual_size);
1060  region_unlock (root_rp);
1061  svm_pop_heap (oldheap);
1062  return;
1063  }
1064 
1065  region_unlock (rp);
1066  region_unlock (root_rp);
1067 
1068  munmap ((void *) virtual_base, virtual_size);
1069 }
1070 
1071 /*
1072  * svm_region_exit
1073  * There is no clean way to unlink the
1074  * root region when all clients go away,
1075  * so remove the pid entry and call it a day.
1076  */
1077 void
1079 {
1080  void *oldheap;
1081  int i, mypid = getpid ();
1082  uword virtual_base, virtual_size;
1083 
1084  /* It felt so nice we did it twice... */
1085  if (root_rp == 0)
1086  return;
1087 
1088  if (--root_rp_refcount > 0)
1089  return;
1090 
1091  /*
1092  * If we take a signal while holding one or more shared-memory
1093  * mutexes, we may end up back here from an otherwise
1094  * benign exit handler. Bail out to avoid a recursive
1095  * mutex screw-up.
1096  */
1097  if (nheld)
1098  return;
1099 
1100  region_lock (root_rp, 7);
1101  oldheap = svm_push_pvt_heap (root_rp);
1102 
1103  virtual_base = root_rp->virtual_base;
1104  virtual_size = root_rp->virtual_size;
1105 
1106  for (i = 0; i < vec_len (root_rp->client_pids); i++)
1107  {
1108  if (root_rp->client_pids[i] == mypid)
1109  {
1110  vec_delete (root_rp->client_pids, 1, i);
1111  goto found;
1112  }
1113  }
1114  clib_warning ("pid %d AWOL", mypid);
1115 
1116 found:
1117 
1118  region_unlock (root_rp);
1119  svm_pop_heap (oldheap);
1120 
1121  root_rp = 0;
1122  munmap ((void *) virtual_base, virtual_size);
1123 }
1124 
1125 void
1127 {
1128  int j;
1129  int mypid = getpid ();
1130  void *oldheap;
1131 
1132  for (j = 0; j < vec_len (rp->client_pids); j++)
1133  {
1134  if (mypid == rp->client_pids[j])
1135  continue;
1136  if (rp->client_pids[j] && (kill (rp->client_pids[j], 0) < 0))
1137  {
1138  clib_warning ("%s: cleanup ghost pid %d",
1139  rp->region_name, rp->client_pids[j]);
1140  /* nb: client vec in rp->region_heap */
1141  oldheap = svm_push_pvt_heap (rp);
1142  vec_delete (rp->client_pids, 1, j);
1143  j--;
1144  svm_pop_heap (oldheap);
1145  }
1146  }
1147 }
1148 
1149 
1150 /*
1151  * Scan svm regions for dead clients
1152  */
1153 void
1154 svm_client_scan (char *root_path)
1155 {
1156  int i, j;
1157  svm_main_region_t *mp;
1158  svm_map_region_args_t *a = 0;
1160  svm_region_t *rp;
1161  svm_subregion_t *subp;
1162  u8 *name = 0;
1163  u8 **svm_names = 0;
1164  void *oldheap;
1165  int mypid = getpid ();
1166 
1167  vec_validate (a, 0);
1168 
1169  svm_region_init_chroot (root_path);
1170 
1171  root_rp = svm_get_root_rp ();
1172 
1173  pthread_mutex_lock (&root_rp->mutex);
1174 
1175  mp = root_rp->data_base;
1176 
1177  for (j = 0; j < vec_len (root_rp->client_pids); j++)
1178  {
1179  if (mypid == root_rp->client_pids[j])
1180  continue;
1181  if (root_rp->client_pids[j] && (kill (root_rp->client_pids[j], 0) < 0))
1182  {
1183  clib_warning ("%s: cleanup ghost pid %d",
1184  root_rp->region_name, root_rp->client_pids[j]);
1185  /* nb: client vec in root_rp->region_heap */
1186  oldheap = svm_push_pvt_heap (root_rp);
1187  vec_delete (root_rp->client_pids, 1, j);
1188  j--;
1189  svm_pop_heap (oldheap);
1190  }
1191  }
1192 
1193  /*
1194  * Snapshoot names, can't hold root rp mutex across
1195  * find_or_create.
1196  */
1197  /* *INDENT-OFF* */
1198  pool_foreach (subp, mp->subregions, ({
1199  name = vec_dup (subp->subregion_name);
1200  vec_add1(svm_names, name);
1201  }));
1202  /* *INDENT-ON* */
1203 
1204  pthread_mutex_unlock (&root_rp->mutex);
1205 
1206  for (i = 0; i < vec_len (svm_names); i++)
1207  {
1208  vec_validate (a, 0);
1209  a->root_path = root_path;
1210  a->name = (char *) svm_names[i];
1211  rp = svm_region_find_or_create (a);
1212  if (rp)
1213  {
1214  pthread_mutex_lock (&rp->mutex);
1215 
1217 
1218  pthread_mutex_unlock (&rp->mutex);
1219  svm_region_unmap (rp);
1220  vec_free (svm_names[i]);
1221  }
1222  vec_free (a);
1223  }
1224  vec_free (svm_names);
1225 
1226  svm_region_exit ();
1227 
1228  vec_free (a);
1229 }
1230 
1231 /*
1232  * fd.io coding-style-patch-verification: ON
1233  *
1234  * Local Variables:
1235  * eval: (c-set-style "gnu")
1236  * End:
1237  */
char * root_path
Definition: svm.h:72
#define vec_validate(V, I)
Make sure vector is long enough for given index (no header, unspecified alignment) ...
Definition: vec.h:396
u8 * root_path
Definition: svm.h:112
vmrglw vmrglh hi
u64 pvt_heap_size
Definition: svm.h:76
svm_region_t * svm_get_root_rp(void)
Definition: svm.c:54
sll srl srl sll sra u16x4 i
Definition: vector_sse2.h:343
static int nheld
Definition: svm.c:51
static void svm_pop_heap(void *oldheap)
Definition: svm.h:190
a
Definition: bitmap.h:516
void * svm_map_region(svm_map_region_args_t *a)
Definition: svm.c:441
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:482
static u64 clib_cpu_time_now(void)
Definition: time.h:73
uword virtual_base
Definition: svm.h:47
#define hash_set_mem(h, key, value)
Definition: hash.h:274
#define pool_get(P, E)
Allocate an object E from a pool P (unspecified alignment).
Definition: pool.h:200
static u8 * format_svm_size(u8 *s, va_list *args)
Definition: svm.c:123
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:1168
#define SVM_FLAGS_MHEAP
Definition: svm.h:32
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
void svm_region_init(void)
Definition: svm.c:782
#define pool_foreach(VAR, POOL, BODY)
Iterate through pool.
Definition: pool.h:348
uword * client_pids
Definition: svm.h:59
#define MHEAP_FLAG_DISABLE_VM
void * svm_region_find_or_create(svm_map_region_args_t *a)
Definition: svm.c:839
volatile void * user_ctx
Definition: svm.h:52
char * name
Definition: svm.h:73
static u64 rnd_pagesize(u64 size)
Definition: svm.c:220
#define clib_warning(format, args...)
Definition: error.h:59
unsigned long u64
Definition: types.h:89
#define SVM_FLAGS_NEED_DATA_INIT
Definition: svm.h:35
#define SVM_GLOBAL_REGION_BASEVA
Definition: svm.h:91
pthread_cond_t condvar
Definition: svm.h:43
u8 * format_svm_region(u8 *s, va_list *args)
Definition: svm.c:143
#define hash_create_string(elts, value_bytes)
Definition: hash.h:652
#define SVM_FLAGS_NODATA
Definition: svm.h:34
void * data_base
Definition: svm.h:50
#define hash_unset_mem(h, key)
Definition: hash.h:280
u8 * subregion_name
Definition: svm.h:105
uword * name_hash
Definition: svm.h:111
#define MAXLOCK
Definition: svm.c:49
#define pool_put(P, E)
Free an object E in pool P.
Definition: pool.h:214
#define vec_dup(V)
Return copy of vector (no header, no alignment)
Definition: vec.h:334
svm_subregion_t * subregions
Definition: svm.h:110
char * backing_file
Definition: svm.h:57
uword virtual_size
Definition: svm.h:48
void svm_region_init_args(svm_map_region_args_t *a)
Definition: svm.c:833
void svm_region_exit()
Definition: svm.c:1078
char * backing_file
Definition: svm.h:78
char * region_name
Definition: svm.h:56
static void * svm_push_pvt_heap(svm_region_t *rp)
Definition: svm.h:174
static int root_rp_refcount
Definition: svm.c:47
static void svm_region_init_internal(svm_map_region_args_t *a)
Definition: svm.c:735
#define vec_free(V)
Free vector&#39;s memory (no header).
Definition: vec.h:300
#define clib_unix_warning(format, args...)
Definition: error.h:68
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:388
uword bitmap_size
Definition: svm.h:54
void svm_region_init_chroot_uid_gid(char *root_path, int uid, int gid)
Definition: svm.c:816
static int svm_data_region_map(svm_map_region_args_t *a, svm_region_t *rp)
Definition: svm.c:314
void * mheap_alloc_with_flags(void *memory, uword memory_size, uword flags)
Definition: mheap.c:875
#define SVM_FLAGS_FILE
Definition: svm.h:33
#define ASSERT(truth)
void svm_region_init_chroot(char *root_path)
Definition: svm.c:799
#define vec_delete(V, N, M)
Delete N elements starting at element M.
Definition: vec.h:745
static void region_lock(svm_region_t *rp, int tag)
Definition: svm.c:62
u32 size
Definition: vhost-user.h:76
volatile uword version
Definition: svm.h:41
Bitmaps built as vectors of machine words.
static void region_unlock(svm_region_t *rp)
Definition: svm.c:79
int mutex_owner_tag
Definition: svm.h:45
#define SVM_GLOBAL_REGION_SIZE
Definition: svm.h:92
#define SVM_VERSION
Definition: svm.h:30
#define MMAP_PAGESIZE
Definition: ssvm.h:41
u64 uword
Definition: types.h:112
#define SVM_GLOBAL_REGION_NAME
Definition: svm.h:93
uword backing_mmap_size
Definition: svm.h:79
#define SVM_PVT_MHEAP_SIZE
Definition: svm.h:37
#define vec_len(v)
Number of elements in vector (rvalue-only, NULL tolerant)
unsigned char u8
Definition: types.h:56
void svm_region_unmap(void *rp_arg)
Definition: svm.c:958
void svm_client_scan_this_region_nolock(svm_region_t *rp)
Definition: svm.c:1126
#define hash_get_mem(h, key)
Definition: hash.h:268
void * region_heap
Definition: svm.h:49
u8 * format(u8 *s, const char *fmt,...)
Definition: format.c:418
uword * bitmap
Definition: svm.h:55
#define CLIB_MEMORY_BARRIER()
Definition: clib.h:101
void * data_heap
Definition: svm.h:51
static int svm_data_region_create(svm_map_region_args_t *a, svm_region_t *rp)
Definition: svm.c:232
static u8 * format_svm_flags(u8 *s, va_list *args)
Definition: svm.c:106
void svm_client_scan(char *root_path)
Definition: svm.c:1154
int mutex_owner_pid
Definition: svm.h:44
#define BITS(x)
Definition: clib.h:58
uword flags
Definition: svm.h:46
static void svm_mutex_cleanup(void)
Definition: svm.c:725
pthread_mutex_t mutex
Definition: svm.h:42
CLIB vectors are ubiquitous dynamically resized arrays with by user defined "headers".
static svm_region_t * root_rp
Definition: svm.c:46