lustre/sec/gss/svcsec_gss.c

   1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
   2  * vim:expandtab:shiftwidth=8:tabstop=8:
   3  *
   4  * Modifications for Lustre
   5  * Copyright 2004, Cluster File Systems, Inc.
   6  * All rights reserved
   7  * Author: Eric Mei <ericm@clusterfs.com>
   8  */
   9
  10 /*
  11  * Neil Brown <neilb@cse.unsw.edu.au>
  12  * J. Bruce Fields <bfields@umich.edu>
  13  * Andy Adamson <andros@umich.edu>
  14  * Dug Song <dugsong@monkey.org>
  15  *
  16  * RPCSEC_GSS server authentication.
  17  * This implements RPCSEC_GSS as defined in rfc2203 (rpcsec_gss) and rfc2078
  18  * (gssapi)
  19  *
  20  * The RPCSEC_GSS involves three stages:
  21  *  1/ context creation
  22  *  2/ data exchange
  23  *  3/ context destruction
  24  *
  25  * Context creation is handled largely by upcalls to user-space.
  26  *  In particular, GSS_Accept_sec_context is handled by an upcall
  27  * Data exchange is handled entirely within the kernel
  28  *  In particular, GSS_GetMIC, GSS_VerifyMIC, GSS_Seal, GSS_Unseal are in-kernel.
  29  * Context destruction is handled in-kernel
  30  *  GSS_Delete_sec_context is in-kernel
  31  *
  32  * Context creation is initiated by a RPCSEC_GSS_INIT request arriving.
  33  * The context handle and gss_token are used as a key into the rpcsec_init cache.
  34  * The content of this cache includes some of the outputs of GSS_Accept_sec_context,
  35  * being major_status, minor_status, context_handle, reply_token.
  36  * These are sent back to the client.
  37  * Sequence window management is handled by the kernel.  The window size if currently
  38  * a compile time constant.
  39  *
  40  * When user-space is happy that a context is established, it places an entry
  41  * in the rpcsec_context cache. The key for this cache is the context_handle.
  42  * The content includes:
  43  *   uid/gidlist - for determining access rights
  44  *   mechanism type
  45  *   mechanism specific information, such as a key
  46  *
  47  */
  48
  49 #define DEBUG_SUBSYSTEM S_SEC
  50 #ifdef __KERNEL__
  51 #include <linux/types.h>
  52 #include <linux/init.h>
  53 #include <linux/module.h>
  54 #include <linux/slab.h>
  55 #include <linux/hash.h>
  56 #else
  57 #include <liblustre.h>
  58 #endif
  59
  60 #include <linux/sunrpc/cache.h>
  61
  62 #include <libcfs/kp30.h>
  63 #include <linux/obd.h>
  64 #include <linux/obd_class.h>
  65 #include <linux/obd_support.h>
  66 #include <linux/lustre_idl.h>
  67 #include <linux/lustre_net.h>
  68 #include <linux/lustre_import.h>
  69 #include <linux/lustre_sec.h>
  70
  71 #include "gss_err.h"
  72 #include "gss_internal.h"
  73 #include "gss_api.h"
  74
  75 static inline unsigned long hash_mem(char *buf, int length, int bits)
  76 {
  77         unsigned long hash = 0;
  78         unsigned long l = 0;
  79         int len = 0;
  80         unsigned char c;
  81         do {
  82                 if (len == length) {
  83                         c = (char)len; len = -1;
  84                 } else
  85                         c = *buf++;
  86                 l = (l << 8) | c;
  87                 len++;
  88                 if ((len & (BITS_PER_LONG/8-1))==0)
  89                         hash = hash_long(hash^l, BITS_PER_LONG);
  90         } while (len);
  91         return hash >> (BITS_PER_LONG - bits);
  92 }
  93
  94 /* The rpcsec_init cache is used for mapping RPCSEC_GSS_{,CONT_}INIT requests
  95  * into replies.
  96  *
  97  * Key is context handle (\x if empty) and gss_token.
  98  * Content is major_status minor_status (integers) context_handle, reply_token.
  99  *
 100  */
 101
 102 #define RSI_HASHBITS    6
 103 #define RSI_HASHMAX     (1<<RSI_HASHBITS)
 104 #define RSI_HASHMASK    (RSI_HASHMAX-1)
 105
 106 struct rsi {
 107         struct cache_head       h;
 108         __u32                   lustre_svc;
 109         __u32                   naltype;
 110         __u32                   netid;
 111         __u64                   nid;
 112         rawobj_t                in_handle, in_token, in_srv_type;
 113         rawobj_t                out_handle, out_token;
 114         int                     major_status, minor_status;
 115 };
 116
 117 static struct cache_head *rsi_table[RSI_HASHMAX];
 118 static struct cache_detail rsi_cache;
 119
 120 static void rsi_free(struct rsi *rsii)
 121 {
 122         rawobj_free(&rsii->in_handle);
 123         rawobj_free(&rsii->in_token);
 124         rawobj_free(&rsii->out_handle);
 125         rawobj_free(&rsii->out_token);
 126 }
 127
 128 static void rsi_put(struct cache_head *item, struct cache_detail *cd)
 129 {
 130         struct rsi *rsii = container_of(item, struct rsi, h);
 131         LASSERT(atomic_read(&item->refcnt) > 0);
 132         if (cache_put(item, cd)) {
 133                 LASSERT(item->next == NULL);
 134                 rsi_free(rsii);
 135                 OBD_FREE(rsii, sizeof(*rsii));
 136         }
 137 }
 138
 139 static inline int rsi_hash(struct rsi *item)
 140 {
 141         return hash_mem((char *)item->in_handle.data, item->in_handle.len, RSI_HASHBITS)
 142                 ^ hash_mem((char *)item->in_token.data, item->in_token.len, RSI_HASHBITS);
 143 }
 144
 145 static inline int rsi_match(struct rsi *item, struct rsi *tmp)
 146 {
 147         return (rawobj_equal(&item->in_handle, &tmp->in_handle) &&
 148                 rawobj_equal(&item->in_token, &tmp->in_token));
 149 }
 150
 151 static void rsi_request(struct cache_detail *cd,
 152                         struct cache_head *h,
 153                         char **bpp, int *blen)
 154 {
 155         struct rsi *rsii = container_of(h, struct rsi, h);
 156
 157         qword_addhex(bpp, blen, (char *) &rsii->lustre_svc,
 158                      sizeof(rsii->lustre_svc));
 159         qword_addhex(bpp, blen, (char *) &rsii->naltype, sizeof(rsii->naltype));
 160         qword_addhex(bpp, blen, (char *) &rsii->netid, sizeof(rsii->netid));
 161         qword_addhex(bpp, blen, (char *) &rsii->nid, sizeof(rsii->nid));
 162         qword_addhex(bpp, blen, rsii->in_handle.data, rsii->in_handle.len);
 163         qword_addhex(bpp, blen, rsii->in_token.data, rsii->in_token.len);
 164         (*bpp)[-1] = '\n';
 165 }
 166
 167 static int
 168 gssd_reply(struct rsi *item)
 169 {
 170         struct rsi *tmp;
 171         struct cache_head **hp, **head;
 172         ENTRY;
 173
 174         head = &rsi_cache.hash_table[rsi_hash(item)];
 175         write_lock(&rsi_cache.hash_lock);
 176         for (hp = head; *hp != NULL; hp = &tmp->h.next) {
 177                 tmp = container_of(*hp, struct rsi, h);
 178                 if (rsi_match(tmp, item)) {
 179                         cache_get(&tmp->h);
 180                         clear_bit(CACHE_HASHED, &tmp->h.flags);
 181                         *hp = tmp->h.next;
 182                         tmp->h.next = NULL;
 183                         rsi_cache.entries--;
 184                         if (test_bit(CACHE_VALID, &tmp->h.flags)) {
 185                                 CERROR("rsi is valid\n");
 186                                 write_unlock(&rsi_cache.hash_lock);
 187                                 rsi_put(&tmp->h, &rsi_cache);
 188                                 RETURN(-EINVAL);
 189                         }
 190                         set_bit(CACHE_HASHED, &item->h.flags);
 191                         item->h.next = *hp;
 192                         *hp = &item->h;
 193                         rsi_cache.entries++;
 194                         set_bit(CACHE_VALID, &item->h.flags);
 195                         item->h.last_refresh = get_seconds();
 196                         write_unlock(&rsi_cache.hash_lock);
 197                         cache_fresh(&rsi_cache, &tmp->h, 0);
 198                         rsi_put(&tmp->h, &rsi_cache);
 199                         RETURN(0);
 200                 }
 201         }
 202         write_unlock(&rsi_cache.hash_lock);
 203         RETURN(-EINVAL);
 204 }
 205
 206 /* XXX
 207  * here we just wait here for its completion or timedout. it's a
 208  * hacking but works, and we'll comeup with real fix if we decided
 209  * to still stick with NFS4 cache code
 210  */
 211 static struct rsi *
 212 gssd_upcall(struct rsi *item, struct cache_req *chandle)
 213 {
 214         struct rsi *tmp;
 215         struct cache_head **hp, **head;
 216         unsigned long starttime;
 217         ENTRY;
 218
 219         head = &rsi_cache.hash_table[rsi_hash(item)];
 220         read_lock(&rsi_cache.hash_lock);
 221         for (hp = head; *hp != NULL; hp = &tmp->h.next) {
 222                 tmp = container_of(*hp, struct rsi, h);
 223                 if (rsi_match(tmp, item)) {
 224                         LBUG();
 225                         if (!test_bit(CACHE_VALID, &tmp->h.flags)) {
 226                                 CERROR("found rsi without VALID\n");
 227                                 read_unlock(&rsi_cache.hash_lock);
 228                                 return NULL;
 229                         }
 230                         *hp = tmp->h.next;
 231                         tmp->h.next = NULL;
 232                         rsi_cache.entries--;
 233                         cache_get(&tmp->h);
 234                         read_unlock(&rsi_cache.hash_lock);
 235                         return tmp;
 236                 }
 237         }
 238         cache_get(&item->h);
 239         set_bit(CACHE_HASHED, &item->h.flags);
 240         item->h.next = *head;
 241         *head = &item->h;
 242         rsi_cache.entries++;
 243         read_unlock(&rsi_cache.hash_lock);
 244         //cache_get(&item->h);
 245
 246         cache_check(&rsi_cache, &item->h, chandle);
 247         starttime = get_seconds();
 248         do {
 249                 set_current_state(TASK_UNINTERRUPTIBLE);
 250                 schedule_timeout(HZ/2);
 251                 read_lock(&rsi_cache.hash_lock);
 252                 for (hp = head; *hp != NULL; hp = &tmp->h.next) {
 253                         tmp = container_of(*hp, struct rsi, h);
 254                         if (tmp == item)
 255                                 continue;
 256                         if (rsi_match(tmp, item)) {
 257                                 if (!test_bit(CACHE_VALID, &tmp->h.flags)) {
 258                                         read_unlock(&rsi_cache.hash_lock);
 259                                         return NULL;
 260                                 }
 261                                 cache_get(&tmp->h);
 262                                 clear_bit(CACHE_HASHED, &tmp->h.flags);
 263                                 *hp = tmp->h.next;
 264                                 tmp->h.next = NULL;
 265                                 rsi_cache.entries--;
 266                                 read_unlock(&rsi_cache.hash_lock);
 267                                 return tmp;
 268                         }
 269                 }
 270                 read_unlock(&rsi_cache.hash_lock);
 271         } while ((get_seconds() - starttime) <= 15);
 272         CERROR("15s timeout while waiting cache refill\n");
 273         return NULL;
 274 }
 275
 276 static int rsi_parse(struct cache_detail *cd,
 277                      char *mesg, int mlen)
 278 {
 279         /* context token expiry major minor context token */
 280         char *buf = mesg;
 281         char *ep;
 282         int len;
 283         struct rsi *rsii;
 284         time_t expiry;
 285         int status = -EINVAL;
 286         ENTRY;
 287
 288         OBD_ALLOC(rsii, sizeof(*rsii));
 289         if (!rsii)
 290                 RETURN(-ENOMEM);
 291         cache_init(&rsii->h);
 292
 293         /* handle */
 294         len = qword_get(&mesg, buf, mlen);
 295         if (len < 0)
 296                 goto out;
 297         if (rawobj_alloc(&rsii->in_handle, buf, len)) {
 298                 status = -ENOMEM;
 299                 goto out;
 300         }
 301
 302         /* token */
 303         len = qword_get(&mesg, buf, mlen);
 304         if (len < 0)
 305                 goto out;
 306         if (rawobj_alloc(&rsii->in_token, buf, len)) {
 307                 status = -ENOMEM;
 308                 goto out;
 309         }
 310
 311         /* expiry */
 312         expiry = get_expiry(&mesg);
 313         if (expiry == 0)
 314                 goto out;
 315
 316         /* major */
 317         len = qword_get(&mesg, buf, mlen);
 318         if (len <= 0)
 319                 goto out;
 320         rsii->major_status = simple_strtol(buf, &ep, 10);
 321         if (*ep)
 322                 goto out;
 323
 324         /* minor */
 325         len = qword_get(&mesg, buf, mlen);
 326         if (len <= 0)
 327                 goto out;
 328         rsii->minor_status = simple_strtol(buf, &ep, 10);
 329         if (*ep)
 330                 goto out;
 331
 332         /* out_handle */
 333         len = qword_get(&mesg, buf, mlen);
 334         if (len < 0)
 335                 goto out;
 336         if (rawobj_alloc(&rsii->out_handle, buf, len)) {
 337                 status = -ENOMEM;
 338                 goto out;
 339         }
 340
 341         /* out_token */
 342         len = qword_get(&mesg, buf, mlen);
 343         if (len < 0)
 344                 goto out;
 345         if (rawobj_alloc(&rsii->out_token, buf, len)) {
 346                 status = -ENOMEM;
 347                 goto out;
 348         }
 349
 350         rsii->h.expiry_time = expiry;
 351         status = gssd_reply(rsii);
 352 out:
 353         if (rsii)
 354                 rsi_put(&rsii->h, &rsi_cache);
 355         RETURN(status);
 356 }
 357
 358 static struct cache_detail rsi_cache = {
 359         .hash_size      = RSI_HASHMAX,
 360         .hash_table     = rsi_table,
 361         .name           = "auth.ptlrpcs.init",
 362         .cache_put      = rsi_put,
 363         .cache_request  = rsi_request,
 364         .cache_parse    = rsi_parse,
 365 };
 366
 367 /*
 368  * The rpcsec_context cache is used to store a context that is
 369  * used in data exchange.
 370  * The key is a context handle. The content is:
 371  *  uid, gidlist, mechanism, service-set, mech-specific-data
 372  */
 373
 374 #define RSC_HASHBITS    10
 375 #define RSC_HASHMAX     (1<<RSC_HASHBITS)
 376 #define RSC_HASHMASK    (RSC_HASHMAX-1)
 377
 378 #define GSS_SEQ_WIN     512
 379
 380 struct gss_svc_seq_data {
 381         /* highest seq number seen so far: */
 382         __u32                   sd_max;
 383         /* for i such that sd_max-GSS_SEQ_WIN < i <= sd_max, the i-th bit of
 384          * sd_win is nonzero iff sequence number i has been seen already: */
 385         unsigned long           sd_win[GSS_SEQ_WIN/BITS_PER_LONG];
 386         spinlock_t              sd_lock;
 387 };
 388
 389 struct rsc {
 390         struct cache_head       h;
 391         rawobj_t                handle;
 392         __u32                   remote_realm:1,
 393                                 auth_usr_mds:1,
 394                                 auth_usr_oss:1;
 395         struct vfs_cred         cred;
 396         uid_t                   mapped_uid;
 397         struct gss_svc_seq_data seqdata;
 398         struct gss_ctx         *mechctx;
 399 };
 400
 401 static struct cache_head *rsc_table[RSC_HASHMAX];
 402 static struct cache_detail rsc_cache;
 403
 404 static void rsc_free(struct rsc *rsci)
 405 {
 406         rawobj_free(&rsci->handle);
 407         if (rsci->mechctx)
 408                 kgss_delete_sec_context(&rsci->mechctx);
 409 #if 0
 410         if (rsci->cred.vc_ginfo)
 411                 put_group_info(rsci->cred.vc_ginfo);
 412 #endif
 413 }
 414
 415 static void rsc_put(struct cache_head *item, struct cache_detail *cd)
 416 {
 417         struct rsc *rsci = container_of(item, struct rsc, h);
 418
 419         LASSERT(atomic_read(&item->refcnt) > 0);
 420         if (cache_put(item, cd)) {
 421                 LASSERT(item->next == NULL);
 422                 rsc_free(rsci);
 423                 OBD_FREE(rsci, sizeof(*rsci));
 424         }
 425 }
 426
 427 static inline int
 428 rsc_hash(struct rsc *rsci)
 429 {
 430         return hash_mem((char *)rsci->handle.data,
 431                         rsci->handle.len, RSC_HASHBITS);
 432 }
 433
 434 static inline int
 435 rsc_match(struct rsc *new, struct rsc *tmp)
 436 {
 437         return rawobj_equal(&new->handle, &tmp->handle);
 438 }
 439
 440 static struct rsc *rsc_lookup(struct rsc *item, int set)
 441 {
 442         struct rsc *tmp = NULL;
 443         struct cache_head **hp, **head;
 444         head = &rsc_cache.hash_table[rsc_hash(item)];
 445         ENTRY;
 446
 447         if (set)
 448                 write_lock(&rsc_cache.hash_lock);
 449         else
 450                 read_lock(&rsc_cache.hash_lock);
 451         for (hp = head; *hp != NULL; hp = &tmp->h.next) {
 452                 tmp = container_of(*hp, struct rsc, h);
 453                 if (!rsc_match(tmp, item))
 454                         continue;
 455                 cache_get(&tmp->h);
 456                 if (!set)
 457                         goto out_noset;
 458                 *hp = tmp->h.next;
 459                 tmp->h.next = NULL;
 460                 clear_bit(CACHE_HASHED, &tmp->h.flags);
 461                 rsc_put(&tmp->h, &rsc_cache);
 462                 goto out_set;
 463         }
 464         /* Didn't find anything */
 465         if (!set)
 466                 goto out_nada;
 467         rsc_cache.entries++;
 468 out_set:
 469         set_bit(CACHE_HASHED, &item->h.flags);
 470         item->h.next = *head;
 471         *head = &item->h;
 472         write_unlock(&rsc_cache.hash_lock);
 473         cache_fresh(&rsc_cache, &item->h, item->h.expiry_time);
 474         cache_get(&item->h);
 475         RETURN(item);
 476 out_nada:
 477         tmp = NULL;
 478 out_noset:
 479         read_unlock(&rsc_cache.hash_lock);
 480         RETURN(tmp);
 481 }
 482
 483 static int rsc_parse(struct cache_detail *cd,
 484                      char *mesg, int mlen)
 485 {
 486         /* contexthandle expiry [ uid gid N <n gids> mechname
 487          * ...mechdata... ] */
 488         char *buf = mesg;
 489         int len, rv, tmp_int;
 490         struct rsc *rsci, *res = NULL;
 491         time_t expiry;
 492         int status = -EINVAL;
 493
 494         OBD_ALLOC(rsci, sizeof(*rsci));
 495         if (!rsci) {
 496                 CERROR("fail to alloc rsci\n");
 497                 return -ENOMEM;
 498         }
 499         cache_init(&rsci->h);
 500
 501         /* context handle */
 502         len = qword_get(&mesg, buf, mlen);
 503         if (len < 0) goto out;
 504         status = -ENOMEM;
 505         if (rawobj_alloc(&rsci->handle, buf, len))
 506                 goto out;
 507
 508         /* expiry */
 509         expiry = get_expiry(&mesg);
 510         status = -EINVAL;
 511         if (expiry == 0)
 512                 goto out;
 513
 514         /* remote flag */
 515         rv = get_int(&mesg, &tmp_int);
 516         if (rv) {
 517                 CERROR("fail to get remote flag\n");
 518                 goto out;
 519         }
 520         rsci->remote_realm = (tmp_int != 0);
 521
 522         /* mds user flag */
 523         rv = get_int(&mesg, &tmp_int);
 524         if (rv) {
 525                 CERROR("fail to get mds user flag\n");
 526                 goto out;
 527         }
 528         rsci->auth_usr_mds = (tmp_int != 0);
 529
 530         /* oss user flag */
 531         rv = get_int(&mesg, &tmp_int);
 532         if (rv) {
 533                 CERROR("fail to get oss user flag\n");
 534                 goto out;
 535         }
 536         rsci->auth_usr_oss = (tmp_int != 0);
 537
 538         /* mapped uid */
 539         rv = get_int(&mesg, (int *)&rsci->mapped_uid);
 540         if (rv) {
 541                 CERROR("fail to get mapped uid\n");
 542                 goto out;
 543         }
 544
 545         /* uid, or NEGATIVE */
 546         rv = get_int(&mesg, (int *)&rsci->cred.vc_uid);
 547         if (rv == -EINVAL)
 548                 goto out;
 549         if (rv == -ENOENT) {
 550                 CERROR("NOENT? set rsc entry negative\n");
 551                 set_bit(CACHE_NEGATIVE, &rsci->h.flags);
 552         } else {
 553                 struct gss_api_mech *gm;
 554                 rawobj_t tmp_buf;
 555                 __u64 ctx_expiry;
 556
 557                 /* gid */
 558                 if (get_int(&mesg, (int *)&rsci->cred.vc_gid))
 559                         goto out;
 560
 561                 /* mech name */
 562                 len = qword_get(&mesg, buf, mlen);
 563                 if (len < 0)
 564                         goto out;
 565                 gm = kgss_name_to_mech(buf);
 566                 status = -EOPNOTSUPP;
 567                 if (!gm)
 568                         goto out;
 569
 570                 status = -EINVAL;
 571                 /* mech-specific data: */
 572                 len = qword_get(&mesg, buf, mlen);
 573                 if (len < 0) {
 574                         kgss_mech_put(gm);
 575                         goto out;
 576                 }
 577                 tmp_buf.len = len;
 578                 tmp_buf.data = (unsigned char *)buf;
 579                 if (kgss_import_sec_context(&tmp_buf, gm, &rsci->mechctx)) {
 580                         kgss_mech_put(gm);
 581                         goto out;
 582                 }
 583
 584                 /* currently the expiry time passed down from user-space
 585                  * is invalid, here we retrive it from mech.
 586                  */
 587                 if (kgss_inquire_context(rsci->mechctx, &ctx_expiry)) {
 588                         CERROR("unable to get expire time, drop it\n");
 589                         set_bit(CACHE_NEGATIVE, &rsci->h.flags);
 590                         kgss_mech_put(gm);
 591                         goto out;
 592                 }
 593                 expiry = (time_t) ctx_expiry;
 594
 595                 kgss_mech_put(gm);
 596         }
 597         rsci->h.expiry_time = expiry;
 598         spin_lock_init(&rsci->seqdata.sd_lock);
 599         res = rsc_lookup(rsci, 1);
 600         rsc_put(&res->h, &rsc_cache);
 601         status = 0;
 602 out:
 603         if (rsci)
 604                 rsc_put(&rsci->h, &rsc_cache);
 605         return status;
 606 }
 607
 608 /*
 609  * flush all entries with @uid. @uid == -1 will match all.
 610  * we only know the uid, maybe netid/nid in the future, in all cases
 611  * we must search the whole cache
 612  */
 613 static void rsc_flush(uid_t uid)
 614 {
 615         struct cache_head **ch;
 616         struct rsc *rscp;
 617         int n;
 618         ENTRY;
 619
 620         write_lock(&rsc_cache.hash_lock);
 621         for (n = 0; n < RSC_HASHMAX; n++) {
 622                 for (ch = &rsc_cache.hash_table[n]; *ch;) {
 623                         rscp = container_of(*ch, struct rsc, h);
 624                         if (uid == -1 || rscp->cred.vc_uid == uid) {
 625                                 /* it seems simply set NEGATIVE doesn't work */
 626                                 *ch = (*ch)->next;
 627                                 rscp->h.next = NULL;
 628                                 cache_get(&rscp->h);
 629                                 set_bit(CACHE_NEGATIVE, &rscp->h.flags);
 630                                 clear_bit(CACHE_HASHED, &rscp->h.flags);
 631                                 CDEBUG(D_SEC, "flush rsc %p for uid %u\n",
 632                                        rscp, rscp->cred.vc_uid);
 633                                 rsc_put(&rscp->h, &rsc_cache);
 634                                 rsc_cache.entries--;
 635                                 continue;
 636                         }
 637                         ch = &((*ch)->next);
 638                 }
 639         }
 640         write_unlock(&rsc_cache.hash_lock);
 641         EXIT;
 642 }
 643
 644 static struct cache_detail rsc_cache = {
 645         .hash_size      = RSC_HASHMAX,
 646         .hash_table     = rsc_table,
 647         .name           = "auth.ptlrpcs.context",
 648         .cache_put      = rsc_put,
 649         .cache_parse    = rsc_parse,
 650 };
 651
 652 static struct rsc *
 653 gss_svc_searchbyctx(rawobj_t *handle)
 654 {
 655         struct rsc rsci;
 656         struct rsc *found;
 657
 658         rsci.handle = *handle;
 659         found = rsc_lookup(&rsci, 0);
 660         if (!found)
 661                 return NULL;
 662
 663         if (cache_check(&rsc_cache, &found->h, NULL))
 664                 return NULL;
 665
 666         return found;
 667 }
 668
 669 /* FIXME
 670  * again hacking: only try to give the svcgssd a chance to handle
 671  * upcalls.
 672  */
 673 struct cache_deferred_req* my_defer(struct cache_req *req)
 674 {
 675         yield();
 676         return NULL;
 677 }
 678 static struct cache_req my_chandle = {my_defer};
 679
 680 /* Implements sequence number algorithm as specified in RFC 2203. */
 681 static int
 682 gss_check_seq_num(struct gss_svc_seq_data *sd, __u32 seq_num)
 683 {
 684         int rc = 0;
 685
 686         spin_lock(&sd->sd_lock);
 687         if (seq_num > sd->sd_max) {
 688                 if (seq_num >= sd->sd_max + GSS_SEQ_WIN) {
 689                         memset(sd->sd_win, 0, sizeof(sd->sd_win));
 690                         sd->sd_max = seq_num;
 691                 } else {
 692                         while(sd->sd_max < seq_num) {
 693                                 sd->sd_max++;
 694                                 __clear_bit(sd->sd_max % GSS_SEQ_WIN,
 695                                             sd->sd_win);
 696                         }
 697                 }
 698                 __set_bit(seq_num % GSS_SEQ_WIN, sd->sd_win);
 699                 goto exit;
 700         } else if (seq_num + GSS_SEQ_WIN <= sd->sd_max) {
 701                 CERROR("seq %u too low: max %u, win %d\n",
 702                         seq_num, sd->sd_max, GSS_SEQ_WIN);
 703                 rc = 1;
 704                 goto exit;
 705         }
 706
 707         if (__test_and_set_bit(seq_num % GSS_SEQ_WIN, sd->sd_win)) {
 708                 CERROR("seq %u is replay: max %u, win %d\n",
 709                         seq_num, sd->sd_max, GSS_SEQ_WIN);
 710                 rc = 1;
 711         }
 712 exit:
 713         spin_unlock(&sd->sd_lock);
 714         return rc;
 715 }
 716
 717 static int
 718 gss_svc_verify_request(struct ptlrpc_request *req,
 719                        struct rsc *rsci,
 720                        struct rpc_gss_wire_cred *gc,
 721                        __u32 *vp, __u32 vlen)
 722 {
 723         struct ptlrpcs_wire_hdr *sec_hdr;
 724         struct gss_ctx *ctx = rsci->mechctx;
 725         __u32 maj_stat;
 726         rawobj_t msg;
 727         rawobj_t mic;
 728         ENTRY;
 729
 730         sec_hdr = (struct ptlrpcs_wire_hdr *) req->rq_reqbuf;
 731
 732         req->rq_reqmsg = (struct lustre_msg *) (req->rq_reqbuf + sizeof(*sec_hdr));
 733         req->rq_reqlen = sec_hdr->msg_len;
 734
 735         msg.len = sec_hdr->msg_len;
 736         msg.data = (__u8 *)req->rq_reqmsg;
 737
 738         mic.len = le32_to_cpu(*vp++);
 739         mic.data = (unsigned char *)vp;
 740         vlen -= 4;
 741
 742         if (mic.len > vlen) {
 743                 CERROR("checksum len %d, while buffer len %d\n",
 744                         mic.len, vlen);
 745                 RETURN(GSS_S_CALL_BAD_STRUCTURE);
 746         }
 747
 748         if (mic.len > 256) {
 749                 CERROR("invalid mic len %d\n", mic.len);
 750                 RETURN(GSS_S_CALL_BAD_STRUCTURE);
 751         }
 752
 753         maj_stat = kgss_verify_mic(ctx, &msg, &mic, NULL);
 754         if (maj_stat != GSS_S_COMPLETE) {
 755                 CERROR("MIC verification error: major %x\n", maj_stat);
 756                 RETURN(maj_stat);
 757         }
 758
 759         if (gss_check_seq_num(&rsci->seqdata, gc->gc_seq)) {
 760                 CERROR("discard replayed request %p(o%u,x"LPU64",t"LPU64")\n",
 761                         req, req->rq_reqmsg->opc, req->rq_xid,
 762                         req->rq_reqmsg->transno);
 763                 RETURN(GSS_S_DUPLICATE_TOKEN);
 764         }
 765
 766         RETURN(GSS_S_COMPLETE);
 767 }
 768
 769 static int
 770 gss_svc_unseal_request(struct ptlrpc_request *req,
 771                        struct rsc *rsci,
 772                        struct rpc_gss_wire_cred *gc,
 773                        __u32 *vp, __u32 vlen)
 774 {
 775         struct ptlrpcs_wire_hdr *sec_hdr;
 776         struct gss_ctx *ctx = rsci->mechctx;
 777         rawobj_t cipher_text, plain_text;
 778         __u32 major;
 779         ENTRY;
 780
 781         sec_hdr = (struct ptlrpcs_wire_hdr *) req->rq_reqbuf;
 782
 783         if (vlen < 4) {
 784                 CERROR("vlen only %u\n", vlen);
 785                 RETURN(GSS_S_CALL_BAD_STRUCTURE);
 786         }
 787
 788         cipher_text.len = le32_to_cpu(*vp++);
 789         cipher_text.data = (__u8 *) vp;
 790         vlen -= 4;
 791
 792         if (cipher_text.len > vlen) {
 793                 CERROR("cipher claimed %u while buf only %u\n",
 794                         cipher_text.len, vlen);
 795                 RETURN(GSS_S_CALL_BAD_STRUCTURE);
 796         }
 797
 798         plain_text = cipher_text;
 799
 800         major = kgss_unwrap(ctx, GSS_C_QOP_DEFAULT, &cipher_text, &plain_text);
 801         if (major) {
 802                 CERROR("unwrap error 0x%x\n", major);
 803                 RETURN(major);
 804         }
 805
 806         if (gss_check_seq_num(&rsci->seqdata, gc->gc_seq)) {
 807                 CERROR("discard replayed request %p(o%u,x"LPU64",t"LPU64")\n",
 808                         req, req->rq_reqmsg->opc, req->rq_xid,
 809                         req->rq_reqmsg->transno);
 810                 RETURN(GSS_S_DUPLICATE_TOKEN);
 811         }
 812
 813         req->rq_reqmsg = (struct lustre_msg *) (vp);
 814         req->rq_reqlen = plain_text.len;
 815
 816         CDEBUG(D_SEC, "msg len %d\n", req->rq_reqlen);
 817
 818         RETURN(GSS_S_COMPLETE);
 819 }
 820
 821 static int
 822 gss_pack_err_notify(struct ptlrpc_request *req,
 823                     __u32 major, __u32 minor)
 824 {
 825         struct gss_svc_data *svcdata = req->rq_svcsec_data;
 826         __u32 reslen, *resp, *reslenp;
 827         char  nidstr[PTL_NALFMT_SIZE];
 828         const __u32 secdata_len = 7 * 4;
 829         int rc;
 830         ENTRY;
 831
 832         OBD_FAIL_RETURN(OBD_FAIL_SVCGSS_ERR_NOTIFY|OBD_FAIL_ONCE, -EINVAL);
 833
 834         LASSERT(svcdata);
 835         svcdata->is_err_notify = 1;
 836         svcdata->reserve_len = 7 * 4;
 837
 838         rc = lustre_pack_reply(req, 0, NULL, NULL);
 839         if (rc) {
 840                 CERROR("could not pack reply, err %d\n", rc);
 841                 RETURN(rc);
 842         }
 843
 844         LASSERT(req->rq_reply_state);
 845         LASSERT(req->rq_reply_state->rs_repbuf);
 846         LASSERT(req->rq_reply_state->rs_repbuf_len >= secdata_len);
 847         resp = (__u32 *) req->rq_reply_state->rs_repbuf;
 848
 849         /* header */
 850         *resp++ = cpu_to_le32(PTLRPCS_FLVR_GSS_NONE);
 851         *resp++ = cpu_to_le32(PTLRPCS_SVC_NONE);
 852         *resp++ = cpu_to_le32(req->rq_replen);
 853         reslenp = resp++;
 854
 855         /* skip lustre msg */
 856         resp += req->rq_replen / 4;
 857         reslen = svcdata->reserve_len;
 858
 859         /* gss replay:
 860          * version, subflavor, notify, major, minor,
 861          * obj1(fake), obj2(fake)
 862          */
 863         *resp++ = cpu_to_le32(PTLRPC_SEC_GSS_VERSION);
 864         *resp++ = cpu_to_le32(PTLRPCS_FLVR_KRB5I);
 865         *resp++ = cpu_to_le32(PTLRPCS_GSS_PROC_ERR);
 866         *resp++ = cpu_to_le32(major);
 867         *resp++ = cpu_to_le32(minor);
 868         *resp++ = 0;
 869         *resp++ = 0;
 870         reslen -= (4 * 4);
 871         /* the actual sec data length */
 872         *reslenp = cpu_to_le32(secdata_len);
 873
 874         req->rq_reply_state->rs_repdata_len += (secdata_len);
 875         CDEBUG(D_SEC, "prepare gss error notify(0x%x/0x%x) to %s\n",
 876                major, minor,
 877                portals_nid2str(req->rq_peer.peer_ni->pni_number,
 878                                req->rq_peer.peer_id.nid, nidstr));
 879         RETURN(0);
 880 }
 881
 882 static void dump_cache_head(struct cache_head *h)
 883 {
 884         CWARN("ref %d, fl %lx, n %p, t %ld, %ld\n",
 885               atomic_read(&h->refcnt), h->flags, h->next,
 886               h->expiry_time, h->last_refresh);
 887 }
 888 static void dump_rsi(struct rsi *rsi)
 889 {
 890         CWARN("dump rsi %p\n", rsi);
 891         dump_cache_head(&rsi->h);
 892         CWARN("%x,%x,%llx\n", rsi->naltype, rsi->netid, rsi->nid);
 893         CWARN("len %d, d %p\n", rsi->in_handle.len, rsi->in_handle.data);
 894         CWARN("len %d, d %p\n", rsi->in_token.len, rsi->in_token.data);
 895         CWARN("len %d, d %p\n", rsi->out_handle.len, rsi->out_handle.data);
 896         CWARN("len %d, d %p\n", rsi->out_token.len, rsi->out_token.data);
 897 }
 898
 899 static int
 900 gss_svcsec_handle_init(struct ptlrpc_request *req,
 901                        struct rpc_gss_wire_cred *gc,
 902                        __u32 *secdata, __u32 seclen,
 903                        enum ptlrpcs_error *res)
 904 {
 905         struct gss_svc_data *svcdata = req->rq_svcsec_data;
 906         struct rsc          *rsci;
 907         struct rsi          *rsikey, *rsip;
 908         rawobj_t             tmpobj;
 909         __u32 reslen,       *resp, *reslenp;
 910         char                 nidstr[PTL_NALFMT_SIZE];
 911         int                  rc;
 912         ENTRY;
 913
 914         LASSERT(svcdata);
 915
 916         CDEBUG(D_SEC, "processing gss init(%d) request from %s\n", gc->gc_proc,
 917                portals_nid2str(req->rq_peer.peer_ni->pni_number,
 918                                req->rq_peer.peer_id.nid, nidstr));
 919
 920         *res = PTLRPCS_BADCRED;
 921         OBD_FAIL_RETURN(OBD_FAIL_SVCGSS_INIT_REQ|OBD_FAIL_ONCE, SVC_DROP);
 922
 923         if (gc->gc_proc == RPC_GSS_PROC_INIT &&
 924             gc->gc_ctx.len != 0) {
 925                 CERROR("proc %d, ctx_len %d: not really init?\n",
 926                 gc->gc_proc == RPC_GSS_PROC_INIT, gc->gc_ctx.len);
 927                 RETURN(SVC_DROP);
 928         }
 929
 930         OBD_ALLOC(rsikey, sizeof(*rsikey));
 931         if (!rsikey) {
 932                 CERROR("out of memory\n");
 933                 RETURN(SVC_DROP);
 934         }
 935         cache_init(&rsikey->h);
 936
 937         /* obtain lustre svc type */
 938         if (seclen < 4) {
 939                 CERROR("sec size %d too small\n", seclen);
 940                 GOTO(out_rsikey, rc = SVC_DROP);
 941         }
 942         rsikey->lustre_svc = le32_to_cpu(*secdata++);
 943         seclen -= 4;
 944
 945         /* duplicate context handle. currently always 0 */
 946         if (rawobj_dup(&rsikey->in_handle, &gc->gc_ctx)) {
 947                 CERROR("fail to dup context handle\n");
 948                 GOTO(out_rsikey, rc = SVC_DROP);
 949         }
 950
 951         /* extract token */
 952         *res = PTLRPCS_BADVERF;
 953         if (rawobj_extract(&tmpobj, &secdata, &seclen)) {
 954                 CERROR("can't extract token\n");
 955                 GOTO(out_rsikey, rc = SVC_DROP);
 956         }
 957         if (rawobj_dup(&rsikey->in_token, &tmpobj)) {
 958                 CERROR("can't duplicate token\n");
 959                 GOTO(out_rsikey, rc = SVC_DROP);
 960         }
 961
 962         rsikey->naltype = (__u32) req->rq_peer.peer_ni->pni_number;
 963         rsikey->netid = 0;
 964         rsikey->nid = (__u64) req->rq_peer.peer_id.nid;
 965
 966         rsip = gssd_upcall(rsikey, &my_chandle);
 967         if (!rsip) {
 968                 CERROR("error in gssd_upcall.\n");
 969
 970                 rc = SVC_COMPLETE;
 971                 if (gss_pack_err_notify(req, GSS_S_FAILURE, 0))
 972                         rc = SVC_DROP;
 973
 974                 GOTO(out_rsikey, rc);
 975         }
 976
 977         rsci = gss_svc_searchbyctx(&rsip->out_handle);
 978         if (!rsci) {
 979                 CERROR("rsci still not mature yet?\n");
 980
 981                 rc = SVC_COMPLETE;
 982                 if (gss_pack_err_notify(req, GSS_S_FAILURE, 0))
 983                         rc = SVC_DROP;
 984
 985                 GOTO(out_rsip, rc);
 986         }
 987         CDEBUG(D_SEC, "svcsec create gss context %p(%u@%s)\n",
 988                rsci, rsci->cred.vc_uid,
 989                portals_nid2str(req->rq_peer.peer_ni->pni_number,
 990                                req->rq_peer.peer_id.nid, nidstr));
 991
 992         svcdata->is_init = 1;
 993         svcdata->reserve_len = 6 * 4 +
 994                 size_round4(rsip->out_handle.len) +
 995                 size_round4(rsip->out_token.len);
 996
 997         rc = lustre_pack_reply(req, 0, NULL, NULL);
 998         if (rc) {
 999                 CERROR("failed to pack reply, rc = %d\n", rc);
1000                 set_bit(CACHE_NEGATIVE, &rsci->h.flags);
1001                 GOTO(out, rc = SVC_DROP);
1002         }
1003
1004         /* header */
1005         resp = (__u32 *) req->rq_reply_state->rs_repbuf;
1006         *resp++ = cpu_to_le32(PTLRPCS_FLVR_GSS_NONE);
1007         *resp++ = cpu_to_le32(PTLRPCS_SVC_NONE);
1008         *resp++ = cpu_to_le32(req->rq_replen);
1009         reslenp = resp++;
1010
1011         resp += req->rq_replen / 4;
1012         reslen = svcdata->reserve_len;
1013
1014         /* gss reply:
1015          * status, major, minor, seq, out_handle, out_token
1016          */
1017         *resp++ = cpu_to_le32(PTLRPCS_OK);
1018         *resp++ = cpu_to_le32(rsip->major_status);
1019         *resp++ = cpu_to_le32(rsip->minor_status);
1020         *resp++ = cpu_to_le32(GSS_SEQ_WIN);
1021         reslen -= (4 * 4);
1022         if (rawobj_serialize(&rsip->out_handle,
1023                              &resp, &reslen)) {
1024                 dump_rsi(rsip);
1025                 dump_rsi(rsikey);
1026                 LBUG();
1027         }
1028         if (rawobj_serialize(&rsip->out_token,
1029                              &resp, &reslen)) {
1030                 dump_rsi(rsip);
1031                 dump_rsi(rsikey);
1032                 LBUG();
1033         }
1034         /* the actual sec data length */
1035         *reslenp = cpu_to_le32(svcdata->reserve_len - reslen);
1036
1037         req->rq_reply_state->rs_repdata_len += le32_to_cpu(*reslenp);
1038         CDEBUG(D_SEC, "req %p: msgsize %d, authsize %d, "
1039                "total size %d\n", req, req->rq_replen,
1040                le32_to_cpu(*reslenp),
1041                req->rq_reply_state->rs_repdata_len);
1042
1043         *res = PTLRPCS_OK;
1044
1045         req->rq_remote_realm = rsci->remote_realm;
1046         req->rq_auth_usr_mds = rsci->auth_usr_mds;
1047         req->rq_auth_usr_oss = rsci->auth_usr_oss;
1048         req->rq_auth_uid = rsci->cred.vc_uid;
1049         req->rq_mapped_uid = rsci->mapped_uid;
1050
1051         if (req->rq_auth_usr_mds) {
1052                 CWARN("usr from %s authenticated as mds svc cred\n",
1053                 portals_nid2str(req->rq_peer.peer_ni->pni_number,
1054                                 req->rq_peer.peer_id.nid, nidstr));
1055         }
1056         if (req->rq_auth_usr_oss) {
1057                 CWARN("usr from %s authenticated as oss svc cred\n",
1058                 portals_nid2str(req->rq_peer.peer_ni->pni_number,
1059                                 req->rq_peer.peer_id.nid, nidstr));
1060         }
1061
1062         /* This is simplified since right now we doesn't support
1063          * INIT_CONTINUE yet.
1064          */
1065         if (gc->gc_proc == RPC_GSS_PROC_INIT) {
1066                 struct ptlrpcs_wire_hdr *hdr;
1067
1068                 hdr = buf_to_sec_hdr(req->rq_reqbuf);
1069                 req->rq_reqmsg = buf_to_lustre_msg(req->rq_reqbuf);
1070                 req->rq_reqlen = hdr->msg_len;
1071
1072                 rc = SVC_LOGIN;
1073         } else
1074                 rc = SVC_COMPLETE;
1075
1076 out:
1077         rsc_put(&rsci->h, &rsc_cache);
1078 out_rsip:
1079         rsi_put(&rsip->h, &rsi_cache);
1080 out_rsikey:
1081         rsi_put(&rsikey->h, &rsi_cache);
1082
1083         RETURN(rc);
1084 }
1085
1086 static int
1087 gss_svcsec_handle_data(struct ptlrpc_request *req,
1088                        struct rpc_gss_wire_cred *gc,
1089                        __u32 *secdata, __u32 seclen,
1090                        enum ptlrpcs_error *res)
1091 {
1092         struct rsc          *rsci;
1093         char                 nidstr[PTL_NALFMT_SIZE];
1094         __u32                major;
1095         int                  rc;
1096         ENTRY;
1097
1098         *res = PTLRPCS_GSS_CREDPROBLEM;
1099
1100         rsci = gss_svc_searchbyctx(&gc->gc_ctx);
1101         if (!rsci) {
1102                 CWARN("Invalid gss context handle from %s\n",
1103                        portals_nid2str(req->rq_peer.peer_ni->pni_number,
1104                                        req->rq_peer.peer_id.nid, nidstr));
1105                 major = GSS_S_NO_CONTEXT;
1106                 goto notify_err;
1107         }
1108
1109         switch (gc->gc_svc) {
1110         case PTLRPCS_GSS_SVC_INTEGRITY:
1111                 major = gss_svc_verify_request(req, rsci, gc, secdata, seclen);
1112                 if (major == GSS_S_COMPLETE)
1113                         break;
1114
1115                 CWARN("fail in verify:0x%x: ctx %p@%s\n", major, rsci,
1116                        portals_nid2str(req->rq_peer.peer_ni->pni_number,
1117                                        req->rq_peer.peer_id.nid, nidstr));
1118                 goto notify_err;
1119         case PTLRPCS_GSS_SVC_PRIVACY:
1120                 major = gss_svc_unseal_request(req, rsci, gc, secdata, seclen);
1121                 if (major == GSS_S_COMPLETE)
1122                         break;
1123
1124                 CWARN("fail in decrypt:0x%x: ctx %p@%s\n", major, rsci,
1125                        portals_nid2str(req->rq_peer.peer_ni->pni_number,
1126                                        req->rq_peer.peer_id.nid, nidstr));
1127                 goto notify_err;
1128         default:
1129                 CERROR("unsupported gss service %d\n", gc->gc_svc);
1130                 GOTO(out, rc = SVC_DROP);
1131         }
1132
1133         req->rq_remote_realm = rsci->remote_realm;
1134         req->rq_auth_usr_mds = rsci->auth_usr_mds;
1135         req->rq_auth_usr_oss = rsci->auth_usr_oss;
1136         req->rq_auth_uid = rsci->cred.vc_uid;
1137         req->rq_mapped_uid = rsci->mapped_uid;
1138
1139         *res = PTLRPCS_OK;
1140         GOTO(out, rc = SVC_OK);
1141
1142 notify_err:
1143         if (gss_pack_err_notify(req, major, 0))
1144                 rc = SVC_DROP;
1145         else
1146                 rc = SVC_COMPLETE;
1147 out:
1148         if (rsci)
1149                 rsc_put(&rsci->h, &rsc_cache);
1150         RETURN(rc);
1151 }
1152
1153 static int
1154 gss_svcsec_handle_destroy(struct ptlrpc_request *req,
1155                           struct rpc_gss_wire_cred *gc,
1156                           __u32 *secdata, __u32 seclen,
1157                           enum ptlrpcs_error *res)
1158 {
1159         struct gss_svc_data *svcdata = req->rq_svcsec_data;
1160         struct rsc          *rsci;
1161         char                 nidstr[PTL_NALFMT_SIZE];
1162         int                  rc;
1163         ENTRY;
1164
1165         LASSERT(svcdata);
1166         *res = PTLRPCS_GSS_CREDPROBLEM;
1167
1168         rsci = gss_svc_searchbyctx(&gc->gc_ctx);
1169         if (!rsci) {
1170                 CWARN("invalid gss context handle for destroy.\n");
1171                 RETURN(SVC_DROP);
1172         }
1173
1174         if (gc->gc_svc != PTLRPCS_GSS_SVC_INTEGRITY) {
1175                 CERROR("service %d is not supported in destroy.\n",
1176                         gc->gc_svc);
1177                 GOTO(out, rc = SVC_DROP);
1178         }
1179
1180         *res = gss_svc_verify_request(req, rsci, gc, secdata, seclen);
1181         if (*res)
1182                 GOTO(out, rc = SVC_DROP);
1183
1184         /* compose reply, which is actually nothing */
1185         svcdata->is_fini = 1;
1186         if (lustre_pack_reply(req, 0, NULL, NULL))
1187                 GOTO(out, rc = SVC_DROP);
1188
1189         CDEBUG(D_SEC, "svcsec destroy gss context %p(%u@%s)\n",
1190                rsci, rsci->cred.vc_uid,
1191                portals_nid2str(req->rq_peer.peer_ni->pni_number,
1192                                req->rq_peer.peer_id.nid, nidstr));
1193
1194         set_bit(CACHE_NEGATIVE, &rsci->h.flags);
1195         *res = PTLRPCS_OK;
1196         rc = SVC_LOGOUT;
1197 out:
1198         rsc_put(&rsci->h, &rsc_cache);
1199         RETURN(rc);
1200 }
1201
1202 /*
1203  * let incomming request go through security check:
1204  *  o context establishment: invoke user space helper
1205  *  o data exchange: verify/decrypt
1206  *  o context destruction: mark context invalid
1207  *
1208  * in most cases, error will result to drop the packet silently.
1209  */
1210 static int
1211 gss_svcsec_accept(struct ptlrpc_request *req, enum ptlrpcs_error *res)
1212 {
1213         struct gss_svc_data *svcdata;
1214         struct rpc_gss_wire_cred *gc;
1215         struct ptlrpcs_wire_hdr *sec_hdr;
1216         __u32 subflavor, seclen, *secdata, version;
1217         int rc;
1218         ENTRY;
1219
1220         CDEBUG(D_SEC, "request %p\n", req);
1221         LASSERT(req->rq_reqbuf);
1222         LASSERT(req->rq_reqbuf_len);
1223
1224         *res = PTLRPCS_BADCRED;
1225
1226         sec_hdr = buf_to_sec_hdr(req->rq_reqbuf);
1227         LASSERT(SEC_FLAVOR_MAJOR(sec_hdr->flavor) == PTLRPCS_FLVR_MAJOR_GSS);
1228
1229         seclen = req->rq_reqbuf_len - sizeof(*sec_hdr) - sec_hdr->msg_len;
1230         secdata = (__u32 *) buf_to_sec_data(req->rq_reqbuf);
1231
1232         if (sec_hdr->sec_len > seclen) {
1233                 CERROR("seclen %d, while max buf %d\n",
1234                         sec_hdr->sec_len, seclen);
1235                 RETURN(SVC_DROP);
1236         }
1237
1238         if (seclen < 6 * 4) {
1239                 CERROR("sec size %d too small\n", seclen);
1240                 RETURN(SVC_DROP);
1241         }
1242
1243         LASSERT(!req->rq_svcsec_data);
1244         OBD_ALLOC(svcdata, sizeof(*svcdata));
1245         if (!svcdata) {
1246                 CERROR("fail to alloc svcdata\n");
1247                 RETURN(SVC_DROP);
1248         }
1249         req->rq_svcsec_data = svcdata;
1250         gc = &svcdata->clcred;
1251
1252         /* Now secdata/seclen is what we want to parse
1253          */
1254         version = le32_to_cpu(*secdata++);      /* version */
1255         subflavor = le32_to_cpu(*secdata++);    /* subflavor */
1256         gc->gc_proc = le32_to_cpu(*secdata++);  /* proc */
1257         gc->gc_seq = le32_to_cpu(*secdata++);   /* seq */
1258         gc->gc_svc = le32_to_cpu(*secdata++);   /* service */
1259         seclen -= 5 * 4;
1260
1261         CDEBUG(D_SEC, "wire gss_hdr: %u/%u/%u/%u/%u\n",
1262                version, subflavor, gc->gc_proc,
1263                gc->gc_seq, gc->gc_svc);
1264
1265         if (version != PTLRPC_SEC_GSS_VERSION) {
1266                 CERROR("gss version mismatch: %d - %d\n",
1267                         version, PTLRPC_SEC_GSS_VERSION);
1268                 GOTO(err_free, rc = SVC_DROP);
1269         }
1270
1271         if (rawobj_extract(&gc->gc_ctx, &secdata, &seclen)) {
1272                 CERROR("fail to obtain gss context handle\n");
1273                 GOTO(err_free, rc = SVC_DROP);
1274         }
1275
1276         *res = PTLRPCS_BADVERF;
1277         switch(gc->gc_proc) {
1278         case RPC_GSS_PROC_INIT:
1279         case RPC_GSS_PROC_CONTINUE_INIT:
1280                 rc = gss_svcsec_handle_init(req, gc, secdata, seclen, res);
1281                 break;
1282         case RPC_GSS_PROC_DATA:
1283                 rc = gss_svcsec_handle_data(req, gc, secdata, seclen, res);
1284                 break;
1285         case RPC_GSS_PROC_DESTROY:
1286                 rc = gss_svcsec_handle_destroy(req, gc, secdata, seclen, res);
1287                 break;
1288         default:
1289                 rc = SVC_DROP;
1290                 LBUG();
1291         }
1292
1293 err_free:
1294         if (rc == SVC_DROP && req->rq_svcsec_data) {
1295                 OBD_FREE(req->rq_svcsec_data, sizeof(struct gss_svc_data));
1296                 req->rq_svcsec_data = NULL;
1297         }
1298
1299         RETURN(rc);
1300 }
1301
1302 static int
1303 gss_svcsec_authorize(struct ptlrpc_request *req)
1304 {
1305         struct ptlrpc_reply_state *rs = req->rq_reply_state;
1306         struct gss_svc_data *gsd = (struct gss_svc_data *)req->rq_svcsec_data;
1307         struct rpc_gss_wire_cred  *gc = &gsd->clcred;
1308         struct rsc                *rscp;
1309         struct ptlrpcs_wire_hdr   *sec_hdr;
1310         rawobj_buf_t               msg_buf;
1311         rawobj_t                   cipher_buf;
1312         __u32                     *vp, *vpsave, major, vlen, seclen;
1313         rawobj_t                   lmsg, mic;
1314         int                        ret;
1315         ENTRY;
1316
1317         LASSERT(rs);
1318         LASSERT(rs->rs_repbuf);
1319         LASSERT(gsd);
1320
1321         if (gsd->is_init || gsd->is_init_continue ||
1322             gsd->is_err_notify || gsd->is_fini) {
1323                 /* nothing to do in these cases */
1324                 CDEBUG(D_SEC, "req %p: init/fini/err\n", req);
1325                 RETURN(0);
1326         }
1327
1328         if (gc->gc_proc != RPC_GSS_PROC_DATA) {
1329                 CERROR("proc %d not support\n", gc->gc_proc);
1330                 RETURN(-EINVAL);
1331         }
1332
1333         rscp = gss_svc_searchbyctx(&gc->gc_ctx);
1334         if (!rscp) {
1335                 CERROR("ctx disapeared under us?\n");
1336                 RETURN(-EINVAL);
1337         }
1338
1339         sec_hdr = (struct ptlrpcs_wire_hdr *) rs->rs_repbuf;
1340         switch (gc->gc_svc) {
1341         case  PTLRPCS_GSS_SVC_INTEGRITY:
1342                 /* prepare various pointers */
1343                 lmsg.len = req->rq_replen;
1344                 lmsg.data = (__u8 *) (rs->rs_repbuf + sizeof(*sec_hdr));
1345                 vp = (__u32 *) (lmsg.data + lmsg.len);
1346                 vlen = rs->rs_repbuf_len - sizeof(*sec_hdr) - lmsg.len;
1347                 seclen = vlen;
1348
1349                 sec_hdr->flavor = cpu_to_le32(PTLRPCS_FLVR_GSS_AUTH);
1350                 sec_hdr->msg_len = cpu_to_le32(req->rq_replen);
1351
1352                 /* standard gss hdr */
1353                 LASSERT(vlen >= 7 * 4);
1354                 *vp++ = cpu_to_le32(PTLRPC_SEC_GSS_VERSION);
1355                 *vp++ = cpu_to_le32(PTLRPCS_FLVR_KRB5I);
1356                 *vp++ = cpu_to_le32(RPC_GSS_PROC_DATA);
1357                 *vp++ = cpu_to_le32(gc->gc_seq);
1358                 *vp++ = cpu_to_le32(PTLRPCS_GSS_SVC_INTEGRITY);
1359                 *vp++ = 0;      /* fake ctx handle */
1360                 vpsave = vp++;  /* reserve size */
1361                 vlen -= 7 * 4;
1362
1363                 mic.len = vlen;
1364                 mic.data = (unsigned char *)vp;
1365
1366                 major = kgss_get_mic(rscp->mechctx, 0, &lmsg, &mic);
1367                 if (major) {
1368                         CERROR("fail to get MIC: 0x%x\n", major);
1369                         GOTO(out, ret = -EINVAL);
1370                 }
1371                 *vpsave = cpu_to_le32(mic.len);
1372                 seclen = seclen - vlen + mic.len;
1373                 sec_hdr->sec_len = cpu_to_le32(seclen);
1374                 rs->rs_repdata_len += size_round(seclen);
1375                 break;
1376         case  PTLRPCS_GSS_SVC_PRIVACY:
1377                 vp = (__u32 *) (rs->rs_repbuf + sizeof(*sec_hdr));
1378                 vlen = rs->rs_repbuf_len - sizeof(*sec_hdr);
1379                 seclen = vlen;
1380
1381                 sec_hdr->flavor = cpu_to_le32(PTLRPCS_FLVR_GSS_PRIV);
1382                 sec_hdr->msg_len = cpu_to_le32(0);
1383
1384                 /* standard gss hdr */
1385                 LASSERT(vlen >= 7 * 4);
1386                 *vp++ = cpu_to_le32(PTLRPC_SEC_GSS_VERSION);
1387                 *vp++ = cpu_to_le32(PTLRPCS_FLVR_KRB5I);
1388                 *vp++ = cpu_to_le32(RPC_GSS_PROC_DATA);
1389                 *vp++ = cpu_to_le32(gc->gc_seq);
1390                 *vp++ = cpu_to_le32(PTLRPCS_GSS_SVC_PRIVACY);
1391                 *vp++ = 0;      /* fake ctx handle */
1392                 vpsave = vp++;  /* reserve size */
1393                 vlen -= 7 * 4;
1394
1395                 msg_buf.buf = (__u8 *) rs->rs_msg - GSS_PRIVBUF_PREFIX_LEN;
1396                 msg_buf.buflen = req->rq_replen + GSS_PRIVBUF_PREFIX_LEN +
1397                                  GSS_PRIVBUF_SUFFIX_LEN;
1398                 msg_buf.dataoff = GSS_PRIVBUF_PREFIX_LEN;
1399                 msg_buf.datalen = req->rq_replen;
1400
1401                 cipher_buf.data = (__u8 *) vp;
1402                 cipher_buf.len = vlen;
1403
1404                 major = kgss_wrap(rscp->mechctx, GSS_C_QOP_DEFAULT,
1405                                 &msg_buf, &cipher_buf);
1406                 if (major) {
1407                         CERROR("failed to wrap: 0x%x\n", major);
1408                         GOTO(out, ret = -EINVAL);
1409                 }
1410
1411                 *vpsave = cpu_to_le32(cipher_buf.len);
1412                 seclen = seclen - vlen + cipher_buf.len;
1413                 sec_hdr->sec_len = cpu_to_le32(seclen);
1414                 rs->rs_repdata_len += size_round(seclen);
1415                 break;
1416         default:
1417                 CERROR("Unknown service %d\n", gc->gc_svc);
1418                 GOTO(out, ret = -EINVAL);
1419         }
1420         ret = 0;
1421 out:
1422         rsc_put(&rscp->h, &rsc_cache);
1423
1424         RETURN(ret);
1425 }
1426
1427 static
1428 void gss_svcsec_cleanup_req(struct ptlrpc_svcsec *svcsec,
1429                             struct ptlrpc_request *req)
1430 {
1431         struct gss_svc_data *gsd = (struct gss_svc_data *) req->rq_svcsec_data;
1432
1433         if (!gsd) {
1434                 CDEBUG(D_SEC, "no svc_data present. do nothing\n");
1435                 return;
1436         }
1437
1438         /* gsd->clclred.gc_ctx is NOT allocated, just set pointer
1439          * to the incoming packet buffer, so don't need free it
1440          */
1441         OBD_FREE(gsd, sizeof(*gsd));
1442         req->rq_svcsec_data = NULL;
1443         return;
1444 }
1445
1446 static
1447 int gss_svcsec_est_payload(struct ptlrpc_svcsec *svcsec,
1448                            struct ptlrpc_request *req,
1449                            int msgsize)
1450 {
1451         struct gss_svc_data *svcdata = req->rq_svcsec_data;
1452         ENTRY;
1453
1454         /* just return the pre-set reserve_len for init/fini/err cases.
1455          */
1456         LASSERT(svcdata);
1457         if (svcdata->is_init) {
1458                 CDEBUG(D_SEC, "is_init, reserver size %d(%d)\n",
1459                        size_round(svcdata->reserve_len),
1460                        svcdata->reserve_len);
1461                 LASSERT(svcdata->reserve_len);
1462                 LASSERT(svcdata->reserve_len % 4 == 0);
1463                 RETURN(size_round(svcdata->reserve_len));
1464         } else if (svcdata->is_err_notify) {
1465                 CDEBUG(D_SEC, "is_err_notify, reserver size %d(%d)\n",
1466                        size_round(svcdata->reserve_len),
1467                        svcdata->reserve_len);
1468                 RETURN(size_round(svcdata->reserve_len));
1469         } else if (svcdata->is_fini) {
1470                 CDEBUG(D_SEC, "is_fini, reserver size 0\n");
1471                 RETURN(0);
1472         } else {
1473                 if (svcdata->clcred.gc_svc == PTLRPCS_GSS_SVC_NONE ||
1474                     svcdata->clcred.gc_svc == PTLRPCS_GSS_SVC_INTEGRITY)
1475                         RETURN(size_round(GSS_MAX_AUTH_PAYLOAD));
1476                 else if (svcdata->clcred.gc_svc == PTLRPCS_GSS_SVC_PRIVACY)
1477                         RETURN(size_round16(GSS_MAX_AUTH_PAYLOAD + msgsize +
1478                                             GSS_PRIVBUF_PREFIX_LEN +
1479                                             GSS_PRIVBUF_SUFFIX_LEN));
1480                 else {
1481                         CERROR("unknown gss svc %u\n", svcdata->clcred.gc_svc);
1482                         *((int *)0) = 0;
1483                         LBUG();
1484                 }
1485         }
1486         RETURN(0);
1487 }
1488
1489 int gss_svcsec_alloc_repbuf(struct ptlrpc_svcsec *svcsec,
1490                             struct ptlrpc_request *req,
1491                             int msgsize)
1492 {
1493         struct gss_svc_data *gsd = (struct gss_svc_data *) req->rq_svcsec_data;
1494         struct ptlrpc_reply_state *rs;
1495         int msg_payload, sec_payload;
1496         int privacy, rc;
1497         ENTRY;
1498
1499         /* determine the security type: none/auth or priv, we have
1500          * different pack scheme for them.
1501          * init/fini/err will always be treated as none/auth.
1502          */
1503         LASSERT(gsd);
1504         if (!gsd->is_init && !gsd->is_init_continue &&
1505             !gsd->is_fini && !gsd->is_err_notify &&
1506             gsd->clcred.gc_svc == PTLRPCS_GSS_SVC_PRIVACY)
1507                 privacy = 1;
1508         else
1509                 privacy = 0;
1510
1511         msg_payload = privacy ? 0 : msgsize;
1512         sec_payload = gss_svcsec_est_payload(svcsec, req, msgsize);
1513
1514         rc = svcsec_alloc_reply_state(req, msg_payload, sec_payload);
1515         if (rc)
1516                 RETURN(rc);
1517
1518         rs = req->rq_reply_state;
1519         LASSERT(rs);
1520         rs->rs_msg_len = msgsize;
1521
1522         if (privacy) {
1523                 /* we can choose to let msg simply point to the rear of the
1524                  * buffer, which lead to buffer overlap when doing encryption.
1525                  * usually it's ok and it indeed passed all existing tests.
1526                  * but not sure if there will be subtle problems in the future.
1527                  * so right now we choose to alloc another new buffer. we'll
1528                  * see how it works.
1529                  */
1530 #if 0
1531                 rs->rs_msg = (struct lustre_msg *)
1532                              (rs->rs_repbuf + rs->rs_repbuf_len -
1533                               msgsize - GSS_PRIVBUF_SUFFIX_LEN);
1534 #endif
1535                 char *msgbuf;
1536
1537                 msgsize += GSS_PRIVBUF_PREFIX_LEN + GSS_PRIVBUF_SUFFIX_LEN;
1538                 OBD_ALLOC(msgbuf, msgsize);
1539                 if (!msgbuf) {
1540                         CERROR("can't alloc %d\n", msgsize);
1541                         svcsec_free_reply_state(rs);
1542                         req->rq_reply_state = NULL;
1543                         RETURN(-ENOMEM);
1544                 }
1545                 rs->rs_msg = (struct lustre_msg *)
1546                                 (msgbuf + GSS_PRIVBUF_PREFIX_LEN);
1547         }
1548
1549         req->rq_repmsg = rs->rs_msg;
1550
1551         RETURN(0);
1552 }
1553
1554 static
1555 void gss_svcsec_free_repbuf(struct ptlrpc_svcsec *svcsec,
1556                             struct ptlrpc_reply_state *rs)
1557 {
1558         unsigned long p1 = (unsigned long) rs->rs_msg;
1559         unsigned long p2 = (unsigned long) rs->rs_buf;
1560
1561         LASSERT(rs->rs_buf);
1562         LASSERT(rs->rs_msg);
1563         LASSERT(rs->rs_msg_len);
1564
1565         if (p1 < p2 || p1 >= p2 + rs->rs_buf_len) {
1566                 char *start = (char*) rs->rs_msg - GSS_PRIVBUF_PREFIX_LEN;
1567                 int size = rs->rs_msg_len + GSS_PRIVBUF_PREFIX_LEN +
1568                            GSS_PRIVBUF_SUFFIX_LEN;
1569                 OBD_FREE(start, size);
1570         }
1571
1572         svcsec_free_reply_state(rs);
1573 }
1574
1575 struct ptlrpc_svcsec svcsec_gss = {
1576         .pss_owner              = THIS_MODULE,
1577         .pss_name               = "svcsec.gss",
1578         .pss_flavor             = PTLRPCS_FLVR_MAJOR_GSS,
1579         .accept                 = gss_svcsec_accept,
1580         .authorize              = gss_svcsec_authorize,
1581         .alloc_repbuf           = gss_svcsec_alloc_repbuf,
1582         .free_repbuf            = gss_svcsec_free_repbuf,
1583         .cleanup_req            = gss_svcsec_cleanup_req,
1584 };
1585
1586 /* XXX hacking */
1587 void lgss_svc_cache_purge_all(void)
1588 {
1589         cache_purge(&rsi_cache);
1590         cache_purge(&rsc_cache);
1591 }
1592 EXPORT_SYMBOL(lgss_svc_cache_purge_all);
1593
1594 void lgss_svc_cache_flush(__u32 uid)
1595 {
1596         rsc_flush(uid);
1597 }
1598 EXPORT_SYMBOL(lgss_svc_cache_flush);
1599
1600 int gss_svc_init(void)
1601 {
1602         int rc;
1603
1604         rc = svcsec_register(&svcsec_gss);
1605         if (!rc) {
1606                 cache_register(&rsc_cache);
1607                 cache_register(&rsi_cache);
1608         }
1609         return rc;
1610 }
1611
1612 void gss_svc_exit(void)
1613 {
1614         int rc;
1615         if ((rc = cache_unregister(&rsi_cache)))
1616                 CERROR("unregister rsi cache: %d\n", rc);
1617         if ((rc = cache_unregister(&rsc_cache)))
1618                 CERROR("unregister rsc cache: %d\n", rc);
1619         if ((rc = svcsec_unregister(&svcsec_gss)))
1620                 CERROR("unregister svcsec_gss: %d\n", rc);
1621 }