FD.io VPP  v20.05.1-6-gf53edbc3b
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 #ifdef CLIB_SANITIZE_ADDR
65  return 0x200000000000;
66 #endif
67 
68 #if __aarch64__
69  /* On AArch64 VA space can have different size, from 36 to 48 bits.
70  Here we are trying to detect VA bits by parsing /proc/self/maps
71  address ranges */
72  int fd;
73  unformat_input_t input;
74  u64 start, end = 0;
75  u8 bits = 0;
76 
77  if ((fd = open ("/proc/self/maps", 0)) < 0)
78  clib_unix_error ("open '/proc/self/maps'");
79 
80  unformat_init_clib_file (&input, fd);
82  {
83  if (unformat (&input, "%llx-%llx", &start, &end))
84  end--;
85  unformat_skip_line (&input);
86  }
87  unformat_free (&input);
88  close (fd);
89 
90  bits = count_leading_zeros (end);
91  bits = 64 - bits;
92  if (bits >= 36 && bits <= 48)
93  return ((1ul << bits) / 4) - (2 * SVM_GLOBAL_REGION_SIZE);
94  else
95  clib_unix_error ("unexpected va bits '%u'", bits);
96 #endif
97 
98  /* default value */
99  return 0x130000000ULL;
100 }
101 
102 static void
103 region_lock (svm_region_t * rp, int tag)
104 {
105  pthread_mutex_lock (&rp->mutex);
106 #ifdef MUTEX_DEBUG
107  rp->mutex_owner_pid = getpid ();
108  rp->mutex_owner_tag = tag;
109 #endif
110  ASSERT (nheld < MAXLOCK); //NOSONAR
111  /*
112  * Keep score of held mutexes so we can try to exit
113  * cleanly if the world comes to an end at the worst possible
114  * moment
115  */
116  mutexes_held[nheld++] = &rp->mutex;
117 }
118 
119 static void
121 {
122  int i, j;
123 #ifdef MUTEX_DEBUG
124  rp->mutex_owner_pid = 0;
125  rp->mutex_owner_tag = 0;
126 #endif
127 
128  for (i = nheld - 1; i >= 0; i--)
129  {
130  if (mutexes_held[i] == &rp->mutex)
131  {
132  for (j = i; j < MAXLOCK - 1; j++)
133  mutexes_held[j] = mutexes_held[j + 1];
134  nheld--;
135  goto found;
136  }
137  }
138  ASSERT (0);
139 
140 found:
142  pthread_mutex_unlock (&rp->mutex);
143 }
144 
145 
146 static u8 *
147 format_svm_flags (u8 * s, va_list * args)
148 {
149  uword f = va_arg (*args, uword);
150 
151  if (f & SVM_FLAGS_MHEAP)
152  s = format (s, "MHEAP ");
153  if (f & SVM_FLAGS_FILE)
154  s = format (s, "FILE ");
155  if (f & SVM_FLAGS_NODATA)
156  s = format (s, "NODATA ");
157  if (f & SVM_FLAGS_NEED_DATA_INIT)
158  s = format (s, "INIT ");
159 
160  return (s);
161 }
162 
163 static u8 *
164 format_svm_size (u8 * s, va_list * args)
165 {
166  uword size = va_arg (*args, uword);
167 
168  if (size >= (1 << 20))
169  {
170  s = format (s, "(%d mb)", size >> 20);
171  }
172  else if (size >= (1 << 10))
173  {
174  s = format (s, "(%d kb)", size >> 10);
175  }
176  else
177  {
178  s = format (s, "(%d bytes)", size);
179  }
180  return (s);
181 }
182 
183 u8 *
184 format_svm_region (u8 * s, va_list * args)
185 {
186  svm_region_t *rp = va_arg (*args, svm_region_t *);
187  int verbose = va_arg (*args, int);
188  int i;
189  uword lo, hi;
190 
191  s = format (s, "%s: base va 0x%x size 0x%x %U\n",
192  rp->region_name, rp->virtual_base,
194  s = format (s, " user_ctx 0x%x, bitmap_size %d\n",
195  rp->user_ctx, rp->bitmap_size);
196 
197  if (verbose)
198  {
199  s = format (s, " flags: 0x%x %U\n", rp->flags,
200  format_svm_flags, rp->flags);
201  s = format (s,
202  " region_heap 0x%x data_base 0x%x data_heap 0x%x\n",
203  rp->region_heap, rp->data_base, rp->data_heap);
204  }
205 
206  s = format (s, " %d clients, pids: ", vec_len (rp->client_pids));
207 
208  for (i = 0; i < vec_len (rp->client_pids); i++)
209  s = format (s, "%d ", rp->client_pids[i]);
210 
211  s = format (s, "\n");
212 
213  if (verbose)
214  {
215  lo = hi = ~0;
216 
217  s = format (s, " VM in use: ");
218 
219  for (i = 0; i < rp->bitmap_size; i++)
220  {
221  if (clib_bitmap_get_no_check (rp->bitmap, i) != 0)
222  {
223  if (lo == ~0)
224  {
225  hi = lo = rp->virtual_base + i * MMAP_PAGESIZE;
226  }
227  else
228  {
229  hi = rp->virtual_base + i * MMAP_PAGESIZE;
230  }
231  }
232  else
233  {
234  if (lo != ~0)
235  {
236  hi = rp->virtual_base + i * MMAP_PAGESIZE - 1;
237  s = format (s, " 0x%x - 0x%x (%dk)\n", lo, hi,
238  (hi - lo) >> 10);
239  lo = hi = ~0;
240  }
241  }
242  }
243  }
244 
245  return (s);
246 }
247 
248 /*
249  * rnd_pagesize
250  * Round to a pagesize multiple, presumably 4k works
251  */
252 static u64
254 {
255  u64 rv;
256 
257  rv = (size + (MMAP_PAGESIZE - 1)) & ~(MMAP_PAGESIZE - 1);
258  return (rv);
259 }
260 
261 /*
262  * svm_data_region_setup
263  */
264 static int
266 {
267  int fd;
268  u8 junk = 0;
269  uword map_size;
270 
271  map_size = rp->virtual_size - (MMAP_PAGESIZE +
272  (a->pvt_heap_size ? a->pvt_heap_size :
274 
275  if (a->flags & SVM_FLAGS_FILE)
276  {
277  struct stat statb;
278 
279  fd = open (a->backing_file, O_RDWR | O_CREAT, 0777);
280 
281  if (fd < 0)
282  {
283  clib_unix_warning ("open");
284  return -1;
285  }
286 
287  if (fstat (fd, &statb) < 0)
288  {
289  clib_unix_warning ("fstat");
290  close (fd);
291  return -2;
292  }
293 
294  if (statb.st_mode & S_IFREG)
295  {
296  if (statb.st_size == 0)
297  {
298  if (lseek (fd, map_size, SEEK_SET) == (off_t) - 1)
299  {
300  clib_unix_warning ("seek region size");
301  close (fd);
302  return -3;
303  }
304  if (write (fd, &junk, 1) != 1)
305  {
306  clib_unix_warning ("set region size");
307  close (fd);
308  return -3;
309  }
310  }
311  else
312  {
313  map_size = rnd_pagesize (statb.st_size);
314  }
315  }
316  else
317  {
318  map_size = a->backing_mmap_size;
319  }
320 
321  ASSERT (map_size <= rp->virtual_size -
323 
324  if (mmap (rp->data_base, map_size, PROT_READ | PROT_WRITE,
325  MAP_SHARED | MAP_FIXED, fd, 0) == MAP_FAILED)
326  {
327  clib_unix_warning ("mmap");
328  close (fd);
329  return -3;
330  }
331  close (fd);
332  CLIB_MEM_UNPOISON (rp->data_base, map_size);
333  rp->backing_file = (char *) format (0, "%s\0", a->backing_file);
334  rp->flags |= SVM_FLAGS_FILE;
335  }
336 
337  if (a->flags & SVM_FLAGS_MHEAP)
338  {
340  map_size, 1 /* locked */ );
342 
343  rp->flags |= SVM_FLAGS_MHEAP;
344  }
345  return 0;
346 }
347 
348 static int
350 {
351  int fd;
352  u8 junk = 0;
353  uword map_size;
354  struct stat statb;
355 
356  map_size = rp->virtual_size -
359 
360  if (a->flags & SVM_FLAGS_FILE)
361  {
362 
363  fd = open (a->backing_file, O_RDWR, 0777);
364 
365  if (fd < 0)
366  {
367  clib_unix_warning ("open");
368  return -1;
369  }
370 
371  if (fstat (fd, &statb) < 0)
372  {
373  clib_unix_warning ("fstat");
374  close (fd);
375  return -2;
376  }
377 
378  if (statb.st_mode & S_IFREG)
379  {
380  if (statb.st_size == 0)
381  {
382  if (lseek (fd, map_size, SEEK_SET) == (off_t) - 1)
383  {
384  clib_unix_warning ("seek region size");
385  close (fd);
386  return -3;
387  }
388  if (write (fd, &junk, 1) != 1)
389  {
390  clib_unix_warning ("set region size");
391  close (fd);
392  return -3;
393  }
394  }
395  else
396  {
397  map_size = rnd_pagesize (statb.st_size);
398  }
399  }
400  else
401  {
402  map_size = a->backing_mmap_size;
403  }
404 
405  ASSERT (map_size <= rp->virtual_size
406  - (MMAP_PAGESIZE
407  +
409 
410  if (mmap (rp->data_base, map_size, PROT_READ | PROT_WRITE,
411  MAP_SHARED | MAP_FIXED, fd, 0) == MAP_FAILED)
412  {
413  clib_unix_warning ("mmap");
414  close (fd);
415  return -3;
416  }
417  close (fd);
418  CLIB_MEM_UNPOISON (rp->data_base, map_size);
419  }
420  return 0;
421 }
422 
423 u8 *
425 {
426  u8 *shm_name;
427  int root_path_offset = 0;
428  int name_offset = 0;
429 
430  if (a->root_path)
431  {
432  /* Tolerate present or absent slashes */
433  if (a->root_path[0] == '/')
434  root_path_offset++;
435 
436  if (a->name[0] == '/')
437  name_offset = 1;
438 
439  shm_name = format (0, "/%s-%s%c", &a->root_path[root_path_offset],
440  &a->name[name_offset], 0);
441  }
442  else
443  shm_name = format (0, "%s%c", a->name, 0);
444  return (shm_name);
445 }
446 
447 void
449 {
450  pthread_mutexattr_t attr;
451  pthread_condattr_t cattr;
452  int nbits, words, bit;
453  int overhead_space;
454  void *oldheap;
455  uword data_base;
456  ASSERT (rp);
457  int rv;
458 
459  clib_memset (rp, 0, sizeof (*rp));
460 
461  if (pthread_mutexattr_init (&attr))
462  clib_unix_warning ("mutexattr_init");
463 
464  if (pthread_mutexattr_setpshared (&attr, PTHREAD_PROCESS_SHARED))
465  clib_unix_warning ("mutexattr_setpshared");
466 
467  if (pthread_mutex_init (&rp->mutex, &attr))
468  clib_unix_warning ("mutex_init");
469 
470  if (pthread_mutexattr_destroy (&attr))
471  clib_unix_warning ("mutexattr_destroy");
472 
473  if (pthread_condattr_init (&cattr))
474  clib_unix_warning ("condattr_init");
475 
476  if (pthread_condattr_setpshared (&cattr, PTHREAD_PROCESS_SHARED))
477  clib_unix_warning ("condattr_setpshared");
478 
479  if (pthread_cond_init (&rp->condvar, &cattr))
480  clib_unix_warning ("cond_init");
481 
482  if (pthread_condattr_destroy (&cattr))
483  clib_unix_warning ("condattr_destroy");
484 
485  region_lock (rp, 1);
486 
487  rp->virtual_base = a->baseva;
488  rp->virtual_size = a->size;
489 
491  (uword_to_pointer (a->baseva + MMAP_PAGESIZE, void *),
492  (a->pvt_heap_size !=
493  0) ? a->pvt_heap_size : SVM_PVT_MHEAP_SIZE, 1 /* locked */ );
494 
496 
497  oldheap = svm_push_pvt_heap (rp);
498 
499  rp->region_name = (char *) format (0, "%s%c", a->name, 0);
500  vec_add1 (rp->client_pids, getpid ());
501 
502  nbits = rp->virtual_size / MMAP_PAGESIZE;
503 
504  ASSERT (nbits > 0);
505  rp->bitmap_size = nbits;
506  words = (nbits + BITS (uword) - 1) / BITS (uword);
507  vec_validate (rp->bitmap, words - 1);
508 
509  overhead_space = MMAP_PAGESIZE /* header */ +
511 
512  bit = 0;
513  data_base = (uword) rp->virtual_base;
514 
515  if (a->flags & SVM_FLAGS_NODATA)
517 
518  do
519  {
520  clib_bitmap_set_no_check (rp->bitmap, bit, 1);
521  bit++;
522  overhead_space -= MMAP_PAGESIZE;
523  data_base += MMAP_PAGESIZE;
524  }
525  while (overhead_space > 0);
526 
527  rp->data_base = (void *) data_base;
528 
529  /*
530  * Note: although the POSIX spec guarantees that only one
531  * process enters this block, we have to play games
532  * to hold off clients until e.g. the mutex is ready
533  */
534  rp->version = SVM_VERSION;
535 
536  /* setup the data portion of the region */
537 
538  rv = svm_data_region_create (a, rp);
539  if (rv)
540  {
541  clib_warning ("data_region_create: %d", rv);
542  }
543 
544  region_unlock (rp);
545 
546  svm_pop_heap (oldheap);
547 }
548 
549 /*
550  * svm_map_region
551  */
552 void *
554 {
555  int svm_fd;
556  svm_region_t *rp;
557  int deadman = 0;
558  u8 junk = 0;
559  void *oldheap;
560  int rv;
561  int pid_holding_region_lock;
562  u8 *shm_name;
563  int dead_region_recovery = 0;
564  int time_left;
565  struct stat stat;
566  struct timespec ts, tsrem;
567 
568  ASSERT ((a->size & ~(MMAP_PAGESIZE - 1)) == a->size);
569  ASSERT (a->name);
570 
571  shm_name = shm_name_from_svm_map_region_args (a);
572 
573  if (CLIB_DEBUG > 1)
574  clib_warning ("[%d] map region %s: shm_open (%s)",
575  getpid (), a->name, shm_name);
576 
577  svm_fd = shm_open ((char *) shm_name, O_RDWR | O_CREAT | O_EXCL, 0777);
578 
579  if (svm_fd >= 0)
580  {
581  if (fchmod (svm_fd, S_IRUSR | S_IWUSR | S_IRGRP | S_IWGRP) < 0)
582  clib_unix_warning ("segment chmod");
583  /* This turns out to fail harmlessly if the client starts first */
584  if (fchown (svm_fd, a->uid, a->gid) < 0)
585  clib_unix_warning ("segment chown [ok if client starts first]");
586 
587  vec_free (shm_name);
588 
589  if (lseek (svm_fd, a->size, SEEK_SET) == (off_t) - 1)
590  {
591  clib_warning ("seek region size");
592  close (svm_fd);
593  return (0);
594  }
595  if (write (svm_fd, &junk, 1) != 1)
596  {
597  clib_warning ("set region size");
598  close (svm_fd);
599  return (0);
600  }
601 
602  rp = mmap (uword_to_pointer (a->baseva, void *), a->size,
603  PROT_READ | PROT_WRITE, MAP_SHARED | MAP_FIXED, svm_fd, 0);
604 
605  if (rp == (svm_region_t *) MAP_FAILED)
606  {
607  clib_unix_warning ("mmap create");
608  close (svm_fd);
609  return (0);
610  }
611  close (svm_fd);
612  CLIB_MEM_UNPOISON (rp, a->size);
613 
615 
616  return ((void *) rp);
617  }
618  else
619  {
620  svm_fd = shm_open ((char *) shm_name, O_RDWR, 0777);
621 
622  vec_free (shm_name);
623 
624  if (svm_fd < 0)
625  {
626  perror ("svm_region_map(mmap open)");
627  return (0);
628  }
629 
630  /* Reset ownership in case the client started first */
631  if (fchown (svm_fd, a->uid, a->gid) < 0)
632  clib_unix_warning ("segment chown [ok if client starts first]");
633 
634  time_left = 20;
635  while (1)
636  {
637  if (0 != fstat (svm_fd, &stat))
638  {
639  clib_warning ("fstat failed: %d", errno);
640  close (svm_fd);
641  return (0);
642  }
643  if (stat.st_size > 0)
644  {
645  break;
646  }
647  if (0 == time_left)
648  {
649  clib_warning ("waiting for resize of shm file timed out");
650  close (svm_fd);
651  return (0);
652  }
653  ts.tv_sec = 0;
654  ts.tv_nsec = 100000000;
655  while (nanosleep (&ts, &tsrem) < 0)
656  ts = tsrem;
657  time_left--;
658  }
659 
660  rp = mmap (0, MMAP_PAGESIZE,
661  PROT_READ | PROT_WRITE, MAP_SHARED, svm_fd, 0);
662 
663  if (rp == (svm_region_t *) MAP_FAILED)
664  {
665  close (svm_fd);
666  clib_warning ("mmap");
667  return (0);
668  }
669 
671 
672  /*
673  * We lost the footrace to create this region; make sure
674  * the winner has crossed the finish line.
675  */
676  while (rp->version == 0 && deadman++ < 5)
677  {
678  sleep (1);
679  }
680 
681  /*
682  * <bleep>-ed?
683  */
684  if (rp->version == 0)
685  {
686  clib_warning ("rp->version %d not %d", rp->version, SVM_VERSION);
687  close (svm_fd);
688  munmap (rp, a->size);
689  return (0);
690  }
691  /* Remap now that the region has been placed */
692  a->baseva = rp->virtual_base;
693  a->size = rp->virtual_size;
694  munmap (rp, MMAP_PAGESIZE);
695 
696  rp = (void *) mmap (uword_to_pointer (a->baseva, void *), a->size,
697  PROT_READ | PROT_WRITE,
698  MAP_SHARED | MAP_FIXED, svm_fd, 0);
699  if ((uword) rp == (uword) MAP_FAILED)
700  {
701  clib_unix_warning ("mmap");
702  close (svm_fd);
703  return (0);
704  }
705 
706  close (svm_fd);
707 
708  CLIB_MEM_UNPOISON (rp, a->size);
709 
710  if ((uword) rp != rp->virtual_base)
711  {
712  clib_warning ("mmap botch");
713  }
714 
715  /*
716  * Try to fix the region mutex if it is held by
717  * a dead process
718  */
719  pid_holding_region_lock = rp->mutex_owner_pid;
720  if (pid_holding_region_lock && kill (pid_holding_region_lock, 0) < 0)
721  {
723  ("region %s mutex held by dead pid %d, tag %d, force unlock",
724  rp->region_name, pid_holding_region_lock, rp->mutex_owner_tag);
725  /* owner pid is nonexistent */
726  rp->mutex.__data.__owner = 0;
727  rp->mutex.__data.__lock = 0;
728  dead_region_recovery = 1;
729  }
730 
731  if (dead_region_recovery)
732  clib_warning ("recovery: attempt to re-lock region");
733 
734  region_lock (rp, 2);
735  oldheap = svm_push_pvt_heap (rp);
736  vec_add1 (rp->client_pids, getpid ());
737 
738  if (dead_region_recovery)
739  clib_warning ("recovery: attempt svm_data_region_map");
740 
741  rv = svm_data_region_map (a, rp);
742  if (rv)
743  {
744  clib_warning ("data_region_map: %d", rv);
745  }
746 
747  if (dead_region_recovery)
748  clib_warning ("unlock and continue");
749 
750  region_unlock (rp);
751 
752  svm_pop_heap (oldheap);
753 
754  return ((void *) rp);
755 
756  }
757  return 0; /* NOTREACHED *///NOSONAR
758 }
759 
760 static void
762 {
763  int i;
764  for (i = 0; i < nheld; i++)
765  {
766  pthread_mutex_unlock (mutexes_held[i]); //NOSONAR
767  }
768 }
769 
770 static int
772 {
773  svm_region_t *rp;
774  u64 ticks = clib_cpu_time_now ();
775  uword randomize_baseva;
776 
777  /* guard against klutz calls */
778  if (root_rp)
779  return -1;
780 
782 
783  atexit (svm_mutex_cleanup);
784 
785  /* Randomize the shared-VM base at init time */
786  if (MMAP_PAGESIZE <= (4 << 10))
787  randomize_baseva = (ticks & 15) * MMAP_PAGESIZE;
788  else
789  randomize_baseva = (ticks & 3) * MMAP_PAGESIZE;
790 
791  a->baseva += randomize_baseva;
792 
793  rp = svm_map_region (a);
794  if (!rp)
795  return -1;
796 
797  region_lock (rp, 3);
798 
799  /* Set up the main region data structures */
801  {
802  svm_main_region_t *mp = 0;
803  void *oldheap;
804 
806 
807  oldheap = svm_push_pvt_heap (rp);
808  vec_validate (mp, 0);
809  mp->name_hash = hash_create_string (0, sizeof (uword));
810  mp->root_path = a->root_path ? format (0, "%s%c", a->root_path, 0) : 0;
811  mp->uid = a->uid;
812  mp->gid = a->gid;
813  rp->data_base = mp;
814  svm_pop_heap (oldheap);
815  }
816  region_unlock (rp);
817  root_rp = rp;
818 
819  return 0;
820 }
821 
822 void
824 {
825  svm_map_region_args_t _a, *a = &_a;
826 
827  clib_memset (a, 0, sizeof (*a));
828  a->root_path = 0;
832  a->flags = SVM_FLAGS_NODATA;
833  a->uid = 0;
834  a->gid = 0;
835 
837 }
838 
839 int
840 svm_region_init_chroot (const char *root_path)
841 {
842  svm_map_region_args_t _a, *a = &_a;
843 
844  clib_memset (a, 0, sizeof (*a));
845  a->root_path = root_path;
849  a->flags = SVM_FLAGS_NODATA;
850  a->uid = 0;
851  a->gid = 0;
852 
853  return svm_region_init_internal (a);
854 }
855 
856 void
857 svm_region_init_chroot_uid_gid (const char *root_path, int uid, int gid)
858 {
859  svm_map_region_args_t _a, *a = &_a;
860 
861  clib_memset (a, 0, sizeof (*a));
862  a->root_path = root_path;
866  a->flags = SVM_FLAGS_NODATA;
867  a->uid = uid;
868  a->gid = gid;
869 
871 }
872 
873 void
875 {
877 }
878 
879 void *
881 {
882  svm_main_region_t *mp;
883  svm_region_t *rp;
884  uword need_nbits;
885  int index, i;
886  void *oldheap;
887  uword *p;
888  u8 *name;
889  svm_subregion_t *subp;
890 
891  ASSERT (root_rp);
892 
893  a->size += MMAP_PAGESIZE +
895  a->size = rnd_pagesize (a->size);
896 
897  region_lock (root_rp, 4);
898  oldheap = svm_push_pvt_heap (root_rp);
899  mp = root_rp->data_base;
900 
901  ASSERT (mp);
902 
903  /* Map the named region from the correct chroot environment */
904  if (a->root_path == NULL)
905  a->root_path = (char *) mp->root_path;
906 
907  /*
908  * See if this region is already known. If it is, we're
909  * almost done...
910  */
911  p = hash_get_mem (mp->name_hash, a->name);
912 
913  if (p)
914  {
915  rp = svm_map_region (a);
916  region_unlock (root_rp);
917  svm_pop_heap (oldheap);
918  return rp;
919  }
920 
921  /* Create the region. */
922  ASSERT ((a->size & ~(MMAP_PAGESIZE - 1)) == a->size);
923 
924  need_nbits = a->size / MMAP_PAGESIZE;
925 
926  index = 1; /* $$$ fixme, figure out how many bit to really skip */
927 
928  /*
929  * Scan the virtual space allocation bitmap, looking for a large
930  * enough chunk
931  */
932  do
933  {
934  if (clib_bitmap_get_no_check (root_rp->bitmap, index) == 0)
935  {
936  for (i = 0; i < (need_nbits - 1); i++)
937  {
938  if (clib_bitmap_get_no_check (root_rp->bitmap, index + i) == 1)
939  {
940  index = index + i;
941  goto next;
942  }
943  }
944  break;
945  }
946  index++;
947  next:;
948  }
949  while (index < root_rp->bitmap_size);
950 
951  /* Completely out of VM? */
952  if (index >= root_rp->bitmap_size)
953  {
954  clib_warning ("region %s: not enough VM to allocate 0x%llx (%lld)",
955  root_rp->region_name, a->size, a->size);
956  svm_pop_heap (oldheap);
957  region_unlock (root_rp);
958  return 0;
959  }
960 
961  /*
962  * Mark virtual space allocated
963  */
964 #if CLIB_DEBUG > 1
965  clib_warning ("set %d bits at index %d", need_nbits, index);
966 #endif
967 
968  for (i = 0; i < need_nbits; i++)
969  {
970  clib_bitmap_set_no_check (root_rp->bitmap, index + i, 1);
971  }
972 
973  /* Place this region where it goes... */
974  a->baseva = root_rp->virtual_base + index * MMAP_PAGESIZE;
975 
976  rp = svm_map_region (a);
977 
978  pool_get (mp->subregions, subp);
979  name = format (0, "%s%c", a->name, 0);
980  subp->subregion_name = name;
981 
982  hash_set_mem (mp->name_hash, name, subp - mp->subregions);
983 
984  svm_pop_heap (oldheap);
985 
986  region_unlock (root_rp);
987 
988  return (rp);
989 }
990 
991 void
993 {
994  svm_map_region_args_t _a, *a = &_a;
995  svm_main_region_t *mp;
996  u8 *shm_name;
997 
998  ASSERT (root_rp);
999  ASSERT (rp);
1001 
1002  mp = root_rp->data_base;
1003  ASSERT (mp);
1004 
1005  a->root_path = (char *) mp->root_path;
1006  a->name = rp->region_name;
1007  shm_name = shm_name_from_svm_map_region_args (a);
1008  if (CLIB_DEBUG > 1)
1009  clib_warning ("[%d] shm_unlink (%s)", getpid (), shm_name);
1010  shm_unlink ((const char *) shm_name);
1011  vec_free (shm_name);
1012 }
1013 
1014 /*
1015  * svm_region_unmap
1016  *
1017  * Let go of the indicated region. If the calling process
1018  * is the last customer, throw it away completely.
1019  * The root region mutex guarantees atomicity with respect to
1020  * a new region client showing up at the wrong moment.
1021  */
1022 void
1023 svm_region_unmap_internal (void *rp_arg, u8 is_client)
1024 {
1025  int i, mypid = getpid ();
1026  int nclients_left;
1027  void *oldheap;
1028  uword virtual_base, virtual_size;
1029  svm_region_t *rp = rp_arg;
1030  char *name;
1031 
1032  /*
1033  * If we take a signal while holding one or more shared-memory
1034  * mutexes, we may end up back here from an otherwise
1035  * benign exit handler. Bail out to avoid a recursive
1036  * mutex screw-up.
1037  */
1038  if (nheld)
1039  return;
1040 
1041  ASSERT (rp);
1042  ASSERT (root_rp);
1043 
1044  if (CLIB_DEBUG > 1)
1045  clib_warning ("[%d] unmap region %s", getpid (), rp->region_name);
1046 
1047  region_lock (root_rp, 5);
1048  region_lock (rp, 6);
1049 
1050  oldheap = svm_push_pvt_heap (rp); /* nb vec_delete() in the loop */
1051 
1052  /* Remove the caller from the list of mappers */
1054  for (i = 0; i < vec_len (rp->client_pids); i++)
1055  {
1056  if (rp->client_pids[i] == mypid)
1057  {
1058  vec_delete (rp->client_pids, 1, i);
1059  goto found;
1060  }
1061  }
1062  clib_warning ("pid %d AWOL", mypid);
1063 
1064 found:
1065 
1066  svm_pop_heap (oldheap);
1067 
1068  nclients_left = vec_len (rp->client_pids);
1069  virtual_base = rp->virtual_base;
1070  virtual_size = rp->virtual_size;
1071 
1072  if (nclients_left == 0)
1073  {
1074  int index, nbits, i;
1075  svm_main_region_t *mp;
1076  uword *p;
1077  svm_subregion_t *subp;
1078 
1079  /* Kill the region, last guy on his way out */
1080 
1081  oldheap = svm_push_pvt_heap (root_rp);
1082  name = vec_dup (rp->region_name);
1083 
1084  virtual_base = rp->virtual_base;
1085  virtual_size = rp->virtual_size;
1086 
1087  /* Figure out which bits to clear in the root region bitmap */
1088  index = (virtual_base - root_rp->virtual_base) / MMAP_PAGESIZE;
1089 
1090  nbits = (virtual_size + MMAP_PAGESIZE - 1) / MMAP_PAGESIZE;
1091 
1092 #if CLIB_DEBUG > 1
1093  clib_warning ("clear %d bits at index %d", nbits, index);
1094 #endif
1095  /* Give back the allocated VM */
1096  for (i = 0; i < nbits; i++)
1097  {
1098  clib_bitmap_set_no_check (root_rp->bitmap, index + i, 0);
1099  }
1100 
1101  mp = root_rp->data_base;
1102 
1103  p = hash_get_mem (mp->name_hash, name);
1104 
1105  /* Better never happen ... */
1106  if (p == NULL)
1107  {
1108  region_unlock (rp);
1109  region_unlock (root_rp);
1110  svm_pop_heap (oldheap);
1111  clib_warning ("Region name '%s' not found?", name);
1112  return;
1113  }
1114 
1115  /* Remove from the root region subregion pool */
1116  subp = mp->subregions + p[0];
1117  pool_put (mp->subregions, subp);
1118 
1119  hash_unset_mem (mp->name_hash, name);
1120 
1121  vec_free (name);
1122 
1123  region_unlock (rp);
1124 
1125  /* If a client asks for the cleanup, don't unlink the backing
1126  * file since we can't tell if it has been recreated. */
1127  if (!is_client)
1128  svm_region_unlink (rp);
1129 
1130  munmap ((void *) virtual_base, virtual_size);
1131  region_unlock (root_rp);
1132  svm_pop_heap (oldheap);
1133  return;
1134  }
1135 
1136  region_unlock (rp);
1137  region_unlock (root_rp);
1138 
1139  munmap ((void *) virtual_base, virtual_size);
1140 }
1141 
1142 void
1143 svm_region_unmap (void *rp_arg)
1144 {
1145  svm_region_unmap_internal (rp_arg, 0 /* is_client */ );
1146 }
1147 
1148 void
1150 {
1151  svm_region_unmap_internal (rp_arg, 1 /* is_client */ );
1152 }
1153 
1154 /*
1155  * svm_region_exit
1156  */
1157 static void
1159 {
1160  void *oldheap;
1161  int i, mypid = getpid ();
1162  uword virtual_base, virtual_size;
1163 
1164  /* It felt so nice we did it twice... */
1165  if (root_rp == 0)
1166  return;
1167 
1168  if (--root_rp_refcount > 0)
1169  return;
1170 
1171  /*
1172  * If we take a signal while holding one or more shared-memory
1173  * mutexes, we may end up back here from an otherwise
1174  * benign exit handler. Bail out to avoid a recursive
1175  * mutex screw-up.
1176  */
1177  if (nheld)
1178  return;
1179 
1180  region_lock (root_rp, 7);
1181  oldheap = svm_push_pvt_heap (root_rp);
1182 
1183  virtual_base = root_rp->virtual_base;
1184  virtual_size = root_rp->virtual_size;
1185 
1186  CLIB_MEM_UNPOISON (root_rp->client_pids, vec_bytes (root_rp->client_pids));
1187  for (i = 0; i < vec_len (root_rp->client_pids); i++)
1188  {
1189  if (root_rp->client_pids[i] == mypid)
1190  {
1191  vec_delete (root_rp->client_pids, 1, i);
1192  goto found;
1193  }
1194  }
1195  clib_warning ("pid %d AWOL", mypid);
1196 
1197 found:
1198 
1199  if (!is_client && vec_len (root_rp->client_pids) == 0)
1200  svm_region_unlink (root_rp);
1201 
1202  region_unlock (root_rp);
1203  svm_pop_heap (oldheap);
1204 
1205  root_rp = 0;
1206  munmap ((void *) virtual_base, virtual_size);
1207 }
1208 
1209 void
1211 {
1212  svm_region_exit_internal (0 /* is_client */ );
1213 }
1214 
1215 void
1217 {
1218  svm_region_exit_internal (1 /* is_client */ );
1219 }
1220 
1221 void
1223 {
1224  int j;
1225  int mypid = getpid ();
1226  void *oldheap;
1227 
1228  for (j = 0; j < vec_len (rp->client_pids); j++)
1229  {
1230  if (mypid == rp->client_pids[j])
1231  continue;
1232  if (rp->client_pids[j] && (kill (rp->client_pids[j], 0) < 0))
1233  {
1234  clib_warning ("%s: cleanup ghost pid %d",
1235  rp->region_name, rp->client_pids[j]);
1236  /* nb: client vec in rp->region_heap */
1237  oldheap = svm_push_pvt_heap (rp);
1238  vec_delete (rp->client_pids, 1, j);
1239  j--;
1240  svm_pop_heap (oldheap);
1241  }
1242  }
1243 }
1244 
1245 
1246 /*
1247  * Scan svm regions for dead clients
1248  */
1249 void
1250 svm_client_scan (const char *root_path)
1251 {
1252  int i, j;
1253  svm_main_region_t *mp;
1254  svm_map_region_args_t *a = 0;
1256  svm_region_t *rp;
1257  svm_subregion_t *subp;
1258  u8 *name = 0;
1259  u8 **svm_names = 0;
1260  void *oldheap;
1261  int mypid = getpid ();
1262 
1263  vec_validate (a, 0);
1264 
1265  svm_region_init_chroot (root_path);
1266 
1267  root_rp = svm_get_root_rp ();
1268 
1269  pthread_mutex_lock (&root_rp->mutex);
1270 
1271  mp = root_rp->data_base;
1272 
1273  for (j = 0; j < vec_len (root_rp->client_pids); j++)
1274  {
1275  if (mypid == root_rp->client_pids[j])
1276  continue;
1277  if (root_rp->client_pids[j] && (kill (root_rp->client_pids[j], 0) < 0))
1278  {
1279  clib_warning ("%s: cleanup ghost pid %d",
1280  root_rp->region_name, root_rp->client_pids[j]);
1281  /* nb: client vec in root_rp->region_heap */
1282  oldheap = svm_push_pvt_heap (root_rp);
1283  vec_delete (root_rp->client_pids, 1, j);
1284  j--;
1285  svm_pop_heap (oldheap);
1286  }
1287  }
1288 
1289  /*
1290  * Snapshoot names, can't hold root rp mutex across
1291  * find_or_create.
1292  */
1293  /* *INDENT-OFF* */
1294  pool_foreach (subp, mp->subregions, ({
1295  name = vec_dup (subp->subregion_name);
1296  vec_add1(svm_names, name);
1297  }));
1298  /* *INDENT-ON* */
1299 
1300  pthread_mutex_unlock (&root_rp->mutex);
1301 
1302  for (i = 0; i < vec_len (svm_names); i++)
1303  {
1304  vec_validate (a, 0);
1305  a->root_path = root_path;
1306  a->name = (char *) svm_names[i];
1307  rp = svm_region_find_or_create (a);
1308  if (rp)
1309  {
1310  pthread_mutex_lock (&rp->mutex);
1311 
1313 
1314  pthread_mutex_unlock (&rp->mutex);
1315  svm_region_unmap (rp);
1316  vec_free (svm_names[i]);
1317  }
1318  vec_free (a);
1319  }
1320  vec_free (svm_names);
1321 
1322  svm_region_exit ();
1323 
1324  vec_free (a);
1325 }
1326 
1327 /*
1328  * fd.io coding-style-patch-verification: ON
1329  *
1330  * Local Variables:
1331  * eval: (c-set-style "gnu")
1332  * End:
1333  */
#define vec_validate(V, I)
Make sure vector is long enough for given index (no header, unspecified alignment) ...
Definition: vec.h:507
void svm_region_init_chroot_uid_gid(const char *root_path, int uid, int gid)
Definition: svm.c:857
#define CLIB_MEM_UNPOISON(a, s)
Definition: sanitizer.h:47
svm_region_t * svm_get_root_rp(void)
Definition: svm.c:54
#define SVM_GLOBAL_REGION_NAME
Definition: svm_common.h:101
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:1080
a
Definition: bitmap.h:538
#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:553
#define count_leading_zeros(x)
Definition: clib.h:153
Optimized string handling code, including c11-compliant "safe C library" variants.
unsigned long u64
Definition: types.h:89
Fixed length block allocator.
clib_memset(h->entries, 0, sizeof(h->entries[0]) *entries)
#define vec_add1(V, E)
Add 1 element to end of vector (unspecified alignment).
Definition: vec.h:590
static u64 clib_cpu_time_now(void)
Definition: time.h:81
void svm_client_scan(const char *root_path)
Definition: svm.c:1250
uword virtual_base
Definition: svm_common.h:42
#define SVM_PVT_MHEAP_SIZE
Definition: svm_common.h:32
void svm_region_unmap_client(void *rp_arg)
Definition: svm.c:1149
#define hash_set_mem(h, key, value)
Definition: hash.h:275
void svm_region_exit_client(void)
Definition: svm.c:1216
u8 * format(u8 *s, const char *fmt,...)
Definition: format.c:424
#define vec_bytes(v)
Number of data bytes in vector.
#define pool_get(P, E)
Allocate an object E from a pool P (unspecified alignment).
Definition: pool.h:252
static u8 * format_svm_size(u8 *s, va_list *args)
Definition: svm.c:164
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
unsigned char u8
Definition: types.h:56
void unformat_init_clib_file(unformat_input_t *input, int file_descriptor)
Definition: unformat.c:1064
#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
DLMALLOC_EXPORT mspace create_mspace_with_base(void *base, size_t capacity, int locked)
void svm_region_exit(void)
Definition: svm.c:1210
static void unformat_skip_line(unformat_input_t *i)
Definition: format.h:222
void svm_region_init(void)
Definition: svm.c:823
#define pool_foreach(VAR, POOL, BODY)
Iterate through pool.
Definition: pool.h:513
uword * client_pids
Definition: svm_common.h:54
void * svm_region_find_or_create(svm_map_region_args_t *a)
Definition: svm.c:880
volatile void * user_ctx
Definition: svm_common.h:47
static u64 rnd_pagesize(u64 size)
Definition: svm.c:253
pthread_cond_t condvar
Definition: svm_common.h:38
u8 * format_svm_region(u8 *s, va_list *args)
Definition: svm.c:184
#define hash_create_string(elts, value_bytes)
Definition: hash.h:690
#define SVM_FLAGS_MHEAP
Definition: svm_common.h:27
static int svm_region_init_internal(svm_map_region_args_t *a)
Definition: svm.c:771
void * data_base
Definition: svm_common.h:45
#define hash_unset_mem(h, key)
Definition: hash.h:291
lo
#define MAXLOCK
Definition: svm.c:49
DLMALLOC_EXPORT void mspace_disable_expand(mspace msp)
struct _unformat_input_t unformat_input_t
u64 size
Definition: vhost_user.h:150
#define pool_put(P, E)
Free an object E in pool P.
Definition: pool.h:302
#define vec_dup(V)
Return copy of vector (no header, no alignment)
Definition: vec.h:427
svm_subregion_t * subregions
Definition: svm_common.h:119
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:874
#define SVM_GLOBAL_REGION_SIZE
Definition: svm_common.h:100
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:448
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:145
sll srl srl sll sra u16x4 i
Definition: vector_sse42.h:317
#define vec_free(V)
Free vector&#39;s memory (no header).
Definition: vec.h:380
#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:424
uword bitmap_size
Definition: svm_common.h:49
string name[64]
Definition: ip.api:44
static int svm_data_region_map(svm_map_region_args_t *a, svm_region_t *rp)
Definition: svm.c:349
#define uword_to_pointer(u, type)
Definition: types.h:136
#define ASSERT(truth)
#define vec_delete(V, N, M)
Delete N elements starting at element M.
Definition: vec.h:852
static void region_lock(svm_region_t *rp, int tag)
Definition: svm.c:103
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:120
int mutex_owner_tag
Definition: svm_common.h:40
void svm_region_unmap_internal(void *rp_arg, u8 is_client)
Definition: svm.c:1023
vl_api_ip4_address_t hi
Definition: arp.api:37
#define MMAP_PAGESIZE
Definition: ssvm.h:43
void svm_region_unlink(svm_region_t *rp)
Definition: svm.c:992
#define vec_len(v)
Number of elements in vector (rvalue-only, NULL tolerant)
const char * name
Definition: svm_common.h:68
u64 uword
Definition: types.h:112
static void unformat_free(unformat_input_t *i)
Definition: format.h:163
#define clib_unix_warning(format, args...)
Definition: error.h:68
void svm_region_unmap(void *rp_arg)
Definition: svm.c:1143
void svm_client_scan_this_region_nolock(svm_region_t *rp)
Definition: svm.c:1222
#define hash_get_mem(h, key)
Definition: hash.h:269
void * region_heap
Definition: svm_common.h:44
uword * bitmap
Definition: svm_common.h:50
f64 end
end of the time range
Definition: mactime.api:44
#define CLIB_MEMORY_BARRIER()
Definition: clib.h:130
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:265
static u8 * format_svm_flags(u8 *s, va_list *args)
Definition: svm.c:147
int mutex_owner_pid
Definition: svm_common.h:39
#define BITS(x)
Definition: clib.h:65
uword flags
Definition: svm_common.h:41
static void svm_mutex_cleanup(void)
Definition: svm.c:761
uword unformat(unformat_input_t *i, const char *fmt,...)
Definition: unformat.c:978
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:171
static svm_region_t * root_rp
Definition: svm.c:46
int svm_region_init_chroot(const char *root_path)
Definition: svm.c:840
static void svm_region_exit_internal(u8 is_client)
Definition: svm.c:1158