kernel/vserver/network.c

   1 /*
   2  *  linux/kernel/vserver/network.c
   3  *
   4  *  Virtual Server: Network Support
   5  *
   6  *  Copyright (C) 2003-2004  Herbert Pötzl
   7  *
   8  *  V0.01  broken out from vcontext V0.05
   9  *  V0.02  cleaned up implementation
  10  *  V0.03  added equiv nx commands
  11  *  V0.04  switch to RCU based hash
  12  *
  13  */
  14
  15 #include <linux/config.h>
  16 #include <linux/slab.h>
  17 #include <linux/vserver.h>
  18 #include <linux/vs_base.h>
  19 #include <linux/vs_network.h>
  20 #include <linux/rcupdate.h>
  21
  22 #include <asm/errno.h>
  23
  24
  25 /*      __alloc_nx_info()
  26
  27         * allocate an initialized nx_info struct
  28         * doesn't make it visible (hash)                        */
  29
  30 static struct nx_info *__alloc_nx_info(nid_t nid)
  31 {
  32         struct nx_info *new = NULL;
  33
  34         vxdprintk(VXD_CBIT(nid, 1), "alloc_nx_info(%d)*", nid);
  35
  36         /* would this benefit from a slab cache? */
  37         new = kmalloc(sizeof(struct nx_info), GFP_KERNEL);
  38         if (!new)
  39                 return 0;
  40
  41         memset (new, 0, sizeof(struct nx_info));
  42         new->nx_id = nid;
  43         INIT_RCU_HEAD(&new->nx_rcu);
  44         INIT_HLIST_NODE(&new->nx_hlist);
  45         atomic_set(&new->nx_refcnt, 0);
  46         atomic_set(&new->nx_usecnt, 0);
  47
  48         /* rest of init goes here */
  49
  50         vxdprintk(VXD_CBIT(nid, 0),
  51                 "alloc_nx_info() = %p", new);
  52         return new;
  53 }
  54
  55 /*      __dealloc_nx_info()
  56
  57         * final disposal of nx_info                             */
  58
  59 static void __dealloc_nx_info(struct nx_info *nxi)
  60 {
  61         vxdprintk(VXD_CBIT(nid, 0),
  62                 "dealloc_nx_info(%p)", nxi);
  63
  64         nxi->nx_hlist.next = LIST_POISON1;
  65         nxi->nx_id = -1;
  66
  67         BUG_ON(atomic_read(&nxi->nx_usecnt));
  68         BUG_ON(atomic_read(&nxi->nx_refcnt));
  69
  70         kfree(nxi);
  71 }
  72
  73
  74 /*      hash table for nx_info hash */
  75
  76 #define NX_HASH_SIZE    13
  77
  78 struct hlist_head nx_info_hash[NX_HASH_SIZE];
  79
  80 static spinlock_t nx_info_hash_lock = SPIN_LOCK_UNLOCKED;
  81
  82
  83 static inline unsigned int __hashval(nid_t nid)
  84 {
  85         return (nid % NX_HASH_SIZE);
  86 }
  87
  88
  89
  90 /*      __hash_nx_info()
  91
  92         * add the nxi to the global hash table
  93         * requires the hash_lock to be held                     */
  94
  95 static inline void __hash_nx_info(struct nx_info *nxi)
  96 {
  97         struct hlist_head *head;
  98
  99         vxdprintk(VXD_CBIT(nid, 4),
 100                 "__hash_nx_info: %p[#%d]", nxi, nxi->nx_id);
 101         get_nx_info(nxi);
 102         head = &nx_info_hash[__hashval(nxi->nx_id)];
 103         hlist_add_head_rcu(&nxi->nx_hlist, head);
 104 }
 105
 106 /*      __unhash_nx_info()
 107
 108         * remove the nxi from the global hash table
 109         * requires the hash_lock to be held                     */
 110
 111 static inline void __unhash_nx_info(struct nx_info *nxi)
 112 {
 113         vxdprintk(VXD_CBIT(nid, 4),
 114                 "__unhash_nx_info: %p[#%d]", nxi, nxi->nx_id);
 115         hlist_del_rcu(&nxi->nx_hlist);
 116         put_nx_info(nxi);
 117 }
 118
 119
 120 /*      __lookup_nx_info()
 121
 122         * requires the rcu_read_lock()
 123         * doesn't increment the nx_refcnt                       */
 124
 125 static inline struct nx_info *__lookup_nx_info(nid_t nid)
 126 {
 127         struct hlist_head *head = &nx_info_hash[__hashval(nid)];
 128         struct hlist_node *pos;
 129
 130         hlist_for_each_rcu(pos, head) {
 131                 struct nx_info *nxi =
 132                         hlist_entry(pos, struct nx_info, nx_hlist);
 133
 134                 if (nxi->nx_id == nid) {
 135                         return nxi;
 136                 }
 137         }
 138         return NULL;
 139 }
 140
 141
 142 /*      __nx_dynamic_id()
 143
 144         * find unused dynamic nid
 145         * requires the hash_lock to be held                     */
 146
 147 static inline nid_t __nx_dynamic_id(void)
 148 {
 149         static nid_t seq = MAX_N_CONTEXT;
 150         nid_t barrier = seq;
 151
 152         do {
 153                 if (++seq > MAX_N_CONTEXT)
 154                         seq = MIN_D_CONTEXT;
 155                 if (!__lookup_nx_info(seq)) {
 156                         vxdprintk(VXD_CBIT(nid, 4),
 157                                 "__nx_dynamic_id: [#%d]", seq);
 158                         return seq;
 159                 }
 160         } while (barrier != seq);
 161         return 0;
 162 }
 163
 164 /*      __loc_nx_info()
 165
 166         * locate or create the requested context
 167         * get() it and if new hash it                           */
 168
 169 static struct nx_info * __loc_nx_info(int id, int *err)
 170 {
 171         struct nx_info *new, *nxi = NULL;
 172
 173         vxdprintk(VXD_CBIT(nid, 1), "loc_nx_info(%d)*", id);
 174
 175         if (!(new = __alloc_nx_info(id))) {
 176                 *err = -ENOMEM;
 177                 return NULL;
 178         }
 179
 180         spin_lock(&nx_info_hash_lock);
 181
 182         /* dynamic context requested */
 183         if (id == NX_DYNAMIC_ID) {
 184                 id = __nx_dynamic_id();
 185                 if (!id) {
 186                         printk(KERN_ERR "no dynamic context available.\n");
 187                         goto out_unlock;
 188                 }
 189                 new->nx_id = id;
 190         }
 191         /* existing context requested */
 192         else if ((nxi = __lookup_nx_info(id))) {
 193                 /* context in setup is not available */
 194                 if (nxi->nx_flags & VXF_STATE_SETUP) {
 195                         vxdprintk(VXD_CBIT(nid, 0),
 196                                 "loc_nx_info(%d) = %p (not available)", id, nxi);
 197                         nxi = NULL;
 198                         *err = -EBUSY;
 199                 } else {
 200                         vxdprintk(VXD_CBIT(nid, 0),
 201                                 "loc_nx_info(%d) = %p (found)", id, nxi);
 202                         get_nx_info(nxi);
 203                         *err = 0;
 204                 }
 205                 goto out_unlock;
 206         }
 207
 208         /* new context requested */
 209         vxdprintk(VXD_CBIT(nid, 0),
 210                 "loc_nx_info(%d) = %p (new)", id, new);
 211         __hash_nx_info(get_nx_info(new));
 212         nxi = new, new = NULL;
 213         *err = 1;
 214
 215 out_unlock:
 216         spin_unlock(&nx_info_hash_lock);
 217         if (new)
 218                 __dealloc_nx_info(new);
 219         return nxi;
 220 }
 221
 222
 223
 224 /*      exported stuff                                          */
 225
 226
 227
 228
 229 void rcu_free_nx_info(struct rcu_head *head)
 230 {
 231         struct nx_info *nxi = container_of(head, struct nx_info, nx_rcu);
 232         int usecnt, refcnt;
 233
 234         BUG_ON(!nxi || !head);
 235
 236         usecnt = atomic_read(&nxi->nx_usecnt);
 237         BUG_ON(usecnt < 0);
 238
 239         refcnt = atomic_read(&nxi->nx_refcnt);
 240         BUG_ON(refcnt < 0);
 241
 242         vxdprintk(VXD_CBIT(nid, 3),
 243                 "rcu_free_nx_info(%p): uc=%d", nxi, usecnt);
 244         if (!usecnt)
 245                 __dealloc_nx_info(nxi);
 246         else
 247                 printk("!!! rcu didn't free\n");
 248 }
 249
 250 void unhash_nx_info(struct nx_info *nxi)
 251 {
 252         spin_lock(&nx_info_hash_lock);
 253         __unhash_nx_info(nxi);
 254         spin_unlock(&nx_info_hash_lock);
 255 }
 256
 257 /*      locate_nx_info()
 258
 259         * search for a nx_info and get() it
 260         * negative id means current                             */
 261
 262 struct nx_info *locate_nx_info(int id)
 263 {
 264         struct nx_info *nxi;
 265
 266         if (id < 0) {
 267                 nxi = get_nx_info(current->nx_info);
 268         } else {
 269                 rcu_read_lock();
 270                 nxi = get_nx_info(__lookup_nx_info(id));
 271                 rcu_read_unlock();
 272         }
 273         return nxi;
 274 }
 275
 276 /*      nx_info_is_hashed()
 277
 278         * verify that nid is still hashed                       */
 279
 280 int nx_info_is_hashed(nid_t nid)
 281 {
 282         int hashed;
 283
 284         rcu_read_lock();
 285         hashed = (__lookup_nx_info(nid) != NULL);
 286         rcu_read_unlock();
 287         return hashed;
 288 }
 289
 290 #ifdef  CONFIG_VSERVER_LEGACY
 291
 292 struct nx_info *locate_or_create_nx_info(int id)
 293 {
 294         int err;
 295
 296         return __loc_nx_info(id, &err);
 297 }
 298
 299 struct nx_info *create_nx_info(void)
 300 {
 301         struct nx_info *new;
 302         int err;
 303
 304         vxdprintk(VXD_CBIT(nid, 5), "create_nx_info(%s)", "void");
 305         if (!(new = __loc_nx_info(NX_DYNAMIC_ID, &err)))
 306                 return NULL;
 307         return new;
 308 }
 309
 310
 311 #endif
 312
 313 #ifdef  CONFIG_PROC_FS
 314
 315 #define hlist_for_each_rcu(pos, head) \
 316         for (pos = (head)->first; pos && ({ prefetch(pos->next); 1;}); \
 317                 pos = pos->next, ({ smp_read_barrier_depends(); 0;}))
 318
 319 int get_nid_list(int index, unsigned int *nids, int size)
 320 {
 321         int hindex, nr_nids = 0;
 322
 323         rcu_read_lock();
 324         for (hindex = 0; hindex < NX_HASH_SIZE; hindex++) {
 325                 struct hlist_head *head = &nx_info_hash[hindex];
 326                 struct hlist_node *pos;
 327
 328                 hlist_for_each_rcu(pos, head) {
 329                         struct nx_info *nxi;
 330
 331                         if (--index > 0)
 332                                 continue;
 333
 334                         nxi = hlist_entry(pos, struct nx_info, nx_hlist);
 335                         nids[nr_nids] = nxi->nx_id;
 336                         if (++nr_nids >= size)
 337                                 goto out;
 338                 }
 339         }
 340 out:
 341         rcu_read_unlock();
 342         return nr_nids;
 343 }
 344 #endif
 345
 346
 347 /*
 348  *      migrate task to new network
 349  */
 350
 351 int nx_migrate_task(struct task_struct *p, struct nx_info *nxi)
 352 {
 353         struct nx_info *old_nxi;
 354         int ret = 0;
 355
 356         if (!p || !nxi)
 357                 BUG();
 358
 359         vxdprintk(VXD_CBIT(nid, 5),
 360                 "nx_migrate_task(%p,%p[#%d.%d.%d])",
 361                 p, nxi, nxi->nx_id,
 362                 atomic_read(&nxi->nx_usecnt),
 363                 atomic_read(&nxi->nx_refcnt));
 364
 365         old_nxi = task_get_nx_info(p);
 366         if (old_nxi == nxi)
 367                 goto out;
 368
 369         task_lock(p);
 370         /* should be handled in set_nx_info !! */
 371         if (old_nxi)
 372                 clr_nx_info(&p->nx_info);
 373         set_nx_info(&p->nx_info, nxi);
 374         p->nid = nxi->nx_id;
 375         task_unlock(p);
 376
 377         /* obsoleted by clr/set */
 378         // put_nx_info(old_nxi);
 379 out:
 380         put_nx_info(old_nxi);
 381         return ret;
 382 }
 383
 384
 385 #include <linux/netdevice.h>
 386 #include <linux/inetdevice.h>
 387
 388 static inline int __addr_in_nx_info(u32 addr, struct nx_info *nxi)
 389 {
 390         int i, nbip;
 391
 392         nbip = nxi->nbipv4;
 393         for (i=0; i<nbip; i++)
 394                 if (nxi->ipv4[i] == addr)
 395                         return 1;
 396         return 0;
 397 }
 398
 399 int ifa_in_nx_info(struct in_ifaddr *ifa, struct nx_info *nxi)
 400 {
 401         if (nxi && ifa)
 402                 return __addr_in_nx_info(ifa->ifa_address, nxi);
 403         return 1;
 404 }
 405
 406 int dev_in_nx_info(struct net_device *dev, struct nx_info *nxi)
 407 {
 408         struct in_device *in_dev = __in_dev_get(dev);
 409         struct in_ifaddr **ifap = NULL;
 410         struct in_ifaddr *ifa = NULL;
 411
 412         if (!nxi)
 413                 return 1;
 414         if (!in_dev)
 415                 return 0;
 416
 417         for (ifap = &in_dev->ifa_list; (ifa = *ifap) != NULL;
 418                 ifap = &ifa->ifa_next) {
 419                 if (__addr_in_nx_info(ifa->ifa_address, nxi))
 420                         return 1;
 421         }
 422         return 0;
 423 }
 424
 425
 426
 427
 428 /* vserver syscall commands below here */
 429
 430 /* taks nid and nx_info functions */
 431
 432 #include <asm/uaccess.h>
 433
 434
 435 int vc_task_nid(uint32_t id, void __user *data)
 436 {
 437         nid_t nid;
 438
 439         if (id) {
 440                 struct task_struct *tsk;
 441
 442                 if (!vx_check(0, VX_ADMIN|VX_WATCH))
 443                         return -EPERM;
 444
 445                 read_lock(&tasklist_lock);
 446                 tsk = find_task_by_pid(id);
 447                 nid = (tsk) ? tsk->nid : -ESRCH;
 448                 read_unlock(&tasklist_lock);
 449         }
 450         else
 451                 nid = current->nid;
 452         return nid;
 453 }
 454
 455
 456 int vc_nx_info(uint32_t id, void __user *data)
 457 {
 458         struct nx_info *nxi;
 459         struct vcmd_nx_info_v0 vc_data;
 460
 461         if (!vx_check(0, VX_ADMIN))
 462                 return -ENOSYS;
 463         if (!capable(CAP_SYS_ADMIN) || !capable(CAP_SYS_RESOURCE))
 464                 return -EPERM;
 465
 466         nxi = locate_nx_info(id);
 467         if (!nxi)
 468                 return -ESRCH;
 469
 470         vc_data.nid = nxi->nx_id;
 471         put_nx_info(nxi);
 472
 473         if (copy_to_user (data, &vc_data, sizeof(vc_data)))
 474                 return -EFAULT;
 475         return 0;
 476 }
 477
 478
 479 /* network functions */
 480
 481 int vc_net_create(uint32_t nid, void __user *data)
 482 {
 483         // int ret = -ENOMEM;
 484         struct nx_info *new_nxi;
 485         int ret;
 486
 487         if (!capable(CAP_SYS_ADMIN))
 488                 return -EPERM;
 489
 490         if ((nid >= MIN_D_CONTEXT) && (nid != VX_DYNAMIC_ID))
 491                 return -EINVAL;
 492
 493         if (nid < 1)
 494                 return -EINVAL;
 495
 496         new_nxi = __loc_nx_info(nid, &ret);
 497         if (!new_nxi)
 498                 return ret;
 499         if (!(new_nxi->nx_flags & VXF_STATE_SETUP)) {
 500                 ret = -EEXIST;
 501                 goto out_put;
 502         }
 503
 504         ret = new_nxi->nx_id;
 505         nx_migrate_task(current, new_nxi);
 506 out_put:
 507         put_nx_info(new_nxi);
 508         return ret;
 509 }
 510
 511
 512 int vc_net_migrate(uint32_t id, void __user *data)
 513 {
 514         struct nx_info *nxi;
 515
 516         if (!capable(CAP_SYS_ADMIN))
 517                 return -EPERM;
 518
 519         nxi = locate_nx_info(id);
 520         if (!nxi)
 521                 return -ESRCH;
 522         nx_migrate_task(current, nxi);
 523         put_nx_info(nxi);
 524         return 0;
 525 }
 526
 527 int vc_net_add(uint32_t id, void __user *data)
 528 {
 529         struct nx_info *nxi;
 530         struct vcmd_net_nx_v0 vc_data;
 531
 532         if (!capable(CAP_SYS_ADMIN))
 533                 return -EPERM;
 534         if (copy_from_user (&vc_data, data, sizeof(vc_data)))
 535                 return -EFAULT;
 536
 537         nxi = locate_nx_info(id);
 538         if (!nxi)
 539                 return -ESRCH;
 540
 541         // add ip to net context here
 542         put_nx_info(nxi);
 543         return 0;
 544 }
 545
 546 int vc_net_remove(uint32_t id, void __user *data)
 547 {
 548         struct nx_info *nxi;
 549         struct vcmd_net_nx_v0 vc_data;
 550
 551         if (!capable(CAP_SYS_ADMIN))
 552                 return -EPERM;
 553         if (copy_from_user (&vc_data, data, sizeof(vc_data)))
 554                 return -EFAULT;
 555
 556         nxi = locate_nx_info(id);
 557         if (!nxi)
 558                 return -ESRCH;
 559
 560         // rem ip from net context here
 561         put_nx_info(nxi);
 562         return 0;
 563 }
 564
 565
 566
 567 int vc_get_nflags(uint32_t id, void __user *data)
 568 {
 569         struct nx_info *nxi;
 570         struct vcmd_net_flags_v0 vc_data;
 571
 572         if (!capable(CAP_SYS_ADMIN))
 573                 return -EPERM;
 574
 575         nxi = locate_nx_info(id);
 576         if (!nxi)
 577                 return -ESRCH;
 578
 579         vc_data.flagword = nxi->nx_flags;
 580
 581         /* special STATE flag handling */
 582         vc_data.mask = vx_mask_flags(~0UL, nxi->nx_flags, IPF_ONE_TIME);
 583
 584         put_nx_info(nxi);
 585
 586         if (copy_to_user (data, &vc_data, sizeof(vc_data)))
 587                 return -EFAULT;
 588         return 0;
 589 }
 590
 591 int vc_set_nflags(uint32_t id, void __user *data)
 592 {
 593         struct nx_info *nxi;
 594         struct vcmd_net_flags_v0 vc_data;
 595         uint64_t mask, trigger;
 596
 597         if (!capable(CAP_SYS_ADMIN))
 598                 return -EPERM;
 599         if (copy_from_user (&vc_data, data, sizeof(vc_data)))
 600                 return -EFAULT;
 601
 602         nxi = locate_nx_info(id);
 603         if (!nxi)
 604                 return -ESRCH;
 605
 606         /* special STATE flag handling */
 607         mask = vx_mask_mask(vc_data.mask, nxi->nx_flags, IPF_ONE_TIME);
 608         trigger = (mask & nxi->nx_flags) ^ (mask & vc_data.flagword);
 609         // if (trigger & IPF_STATE_SETUP)
 610
 611         nxi->nx_flags = vx_mask_flags(nxi->nx_flags,
 612                 vc_data.flagword, mask);
 613         put_nx_info(nxi);
 614         return 0;
 615 }
 616
 617 int vc_get_ncaps(uint32_t id, void __user *data)
 618 {
 619         struct nx_info *nxi;
 620         struct vcmd_net_caps_v0 vc_data;
 621
 622         if (!capable(CAP_SYS_ADMIN))
 623                 return -EPERM;
 624
 625         nxi = locate_nx_info(id);
 626         if (!nxi)
 627                 return -ESRCH;
 628
 629         vc_data.ncaps = nxi->nx_ncaps;
 630         vc_data.cmask = ~0UL;
 631         put_nx_info(nxi);
 632
 633         if (copy_to_user (data, &vc_data, sizeof(vc_data)))
 634                 return -EFAULT;
 635         return 0;
 636 }
 637
 638 int vc_set_ncaps(uint32_t id, void __user *data)
 639 {
 640         struct nx_info *nxi;
 641         struct vcmd_net_caps_v0 vc_data;
 642
 643         if (!capable(CAP_SYS_ADMIN))
 644                 return -EPERM;
 645         if (copy_from_user (&vc_data, data, sizeof(vc_data)))
 646                 return -EFAULT;
 647
 648         nxi = locate_nx_info(id);
 649         if (!nxi)
 650                 return -ESRCH;
 651
 652         nxi->nx_ncaps = vx_mask_flags(nxi->nx_ncaps,
 653                 vc_data.ncaps, vc_data.cmask);
 654         put_nx_info(nxi);
 655         return 0;
 656 }
 657
 658
 659 #include <linux/module.h>
 660
 661 EXPORT_SYMBOL_GPL(rcu_free_nx_info);
 662 EXPORT_SYMBOL_GPL(nx_info_hash_lock);
 663 EXPORT_SYMBOL_GPL(unhash_nx_info);
 664