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) <= SVCSEC_UPCALL_TIMEOUT);
 272         CERROR("%ds timeout while waiting cache refill\n",
 273                SVCSEC_UPCALL_TIMEOUT);
 274         return NULL;
 275 }
 276
 277 static int rsi_parse(struct cache_detail *cd,
 278                      char *mesg, int mlen)
 279 {
 280         /* context token expiry major minor context token */
 281         char *buf = mesg;
 282         char *ep;
 283         int len;
 284         struct rsi *rsii;
 285         time_t expiry;
 286         int status = -EINVAL;
 287         ENTRY;
 288
 289         OBD_ALLOC(rsii, sizeof(*rsii));
 290         if (!rsii)
 291                 RETURN(-ENOMEM);
 292         cache_init(&rsii->h);
 293
 294         /* handle */
 295         len = qword_get(&mesg, buf, mlen);
 296         if (len < 0)
 297                 goto out;
 298         if (rawobj_alloc(&rsii->in_handle, buf, len)) {
 299                 status = -ENOMEM;
 300                 goto out;
 301         }
 302
 303         /* token */
 304         len = qword_get(&mesg, buf, mlen);
 305         if (len < 0)
 306                 goto out;
 307         if (rawobj_alloc(&rsii->in_token, buf, len)) {
 308                 status = -ENOMEM;
 309                 goto out;
 310         }
 311
 312         /* expiry */
 313         expiry = get_expiry(&mesg);
 314         if (expiry == 0)
 315                 goto out;
 316
 317         /* major */
 318         len = qword_get(&mesg, buf, mlen);
 319         if (len <= 0)
 320                 goto out;
 321         rsii->major_status = simple_strtol(buf, &ep, 10);
 322         if (*ep)
 323                 goto out;
 324
 325         /* minor */
 326         len = qword_get(&mesg, buf, mlen);
 327         if (len <= 0)
 328                 goto out;
 329         rsii->minor_status = simple_strtol(buf, &ep, 10);
 330         if (*ep)
 331                 goto out;
 332
 333         /* out_handle */
 334         len = qword_get(&mesg, buf, mlen);
 335         if (len < 0)
 336                 goto out;
 337         if (rawobj_alloc(&rsii->out_handle, buf, len)) {
 338                 status = -ENOMEM;
 339                 goto out;
 340         }
 341
 342         /* out_token */
 343         len = qword_get(&mesg, buf, mlen);
 344         if (len < 0)
 345                 goto out;
 346         if (rawobj_alloc(&rsii->out_token, buf, len)) {
 347                 status = -ENOMEM;
 348                 goto out;
 349         }
 350
 351         rsii->h.expiry_time = expiry;
 352         status = gssd_reply(rsii);
 353 out:
 354         if (rsii)
 355                 rsi_put(&rsii->h, &rsi_cache);
 356         RETURN(status);
 357 }
 358
 359 static struct cache_detail rsi_cache = {
 360         .hash_size      = RSI_HASHMAX,
 361         .hash_table     = rsi_table,
 362         .name           = "auth.ptlrpcs.init",
 363         .cache_put      = rsi_put,
 364         .cache_request  = rsi_request,
 365         .cache_parse    = rsi_parse,
 366 };
 367
 368 /*
 369  * The rpcsec_context cache is used to store a context that is
 370  * used in data exchange.
 371  * The key is a context handle. The content is:
 372  *  uid, gidlist, mechanism, service-set, mech-specific-data
 373  */
 374
 375 #define RSC_HASHBITS    10
 376 #define RSC_HASHMAX     (1<<RSC_HASHBITS)
 377 #define RSC_HASHMASK    (RSC_HASHMAX-1)
 378
 379 #define GSS_SEQ_WIN     512
 380
 381 struct gss_svc_seq_data {
 382         /* highest seq number seen so far: */
 383         __u32                   sd_max;
 384         /* for i such that sd_max-GSS_SEQ_WIN < i <= sd_max, the i-th bit of
 385          * sd_win is nonzero iff sequence number i has been seen already: */
 386         unsigned long           sd_win[GSS_SEQ_WIN/BITS_PER_LONG];
 387         spinlock_t              sd_lock;
 388 };
 389
 390 struct rsc {
 391         struct cache_head       h;
 392         rawobj_t                handle;
 393         __u32                   remote_realm:1,
 394                                 auth_usr_mds:1,
 395                                 auth_usr_oss:1;
 396         struct vfs_cred         cred;
 397         uid_t                   mapped_uid;
 398         struct gss_svc_seq_data seqdata;
 399         struct gss_ctx         *mechctx;
 400 };
 401
 402 static struct cache_head *rsc_table[RSC_HASHMAX];
 403 static struct cache_detail rsc_cache;
 404
 405 static void rsc_free(struct rsc *rsci)
 406 {
 407         rawobj_free(&rsci->handle);
 408         if (rsci->mechctx)
 409                 kgss_delete_sec_context(&rsci->mechctx);
 410 #if 0
 411         if (rsci->cred.vc_ginfo)
 412                 put_group_info(rsci->cred.vc_ginfo);
 413 #endif
 414 }
 415
 416 static void rsc_put(struct cache_head *item, struct cache_detail *cd)
 417 {
 418         struct rsc *rsci = container_of(item, struct rsc, h);
 419
 420         LASSERT(atomic_read(&item->refcnt) > 0);
 421         if (cache_put(item, cd)) {
 422                 LASSERT(item->next == NULL);
 423                 rsc_free(rsci);
 424                 OBD_FREE(rsci, sizeof(*rsci));
 425         }
 426 }
 427
 428 static inline int
 429 rsc_hash(struct rsc *rsci)
 430 {
 431         return hash_mem((char *)rsci->handle.data,
 432                         rsci->handle.len, RSC_HASHBITS);
 433 }
 434
 435 static inline int
 436 rsc_match(struct rsc *new, struct rsc *tmp)
 437 {
 438         return rawobj_equal(&new->handle, &tmp->handle);
 439 }
 440
 441 static struct rsc *rsc_lookup(struct rsc *item, int set)
 442 {
 443         struct rsc *tmp = NULL;
 444         struct cache_head **hp, **head;
 445         head = &rsc_cache.hash_table[rsc_hash(item)];
 446         ENTRY;
 447
 448         if (set)
 449                 write_lock(&rsc_cache.hash_lock);
 450         else
 451                 read_lock(&rsc_cache.hash_lock);
 452         for (hp = head; *hp != NULL; hp = &tmp->h.next) {
 453                 tmp = container_of(*hp, struct rsc, h);
 454                 if (!rsc_match(tmp, item))
 455                         continue;
 456                 cache_get(&tmp->h);
 457                 if (!set)
 458                         goto out_noset;
 459                 *hp = tmp->h.next;
 460                 tmp->h.next = NULL;
 461                 clear_bit(CACHE_HASHED, &tmp->h.flags);
 462                 rsc_put(&tmp->h, &rsc_cache);
 463                 goto out_set;
 464         }
 465         /* Didn't find anything */
 466         if (!set)
 467                 goto out_nada;
 468         rsc_cache.entries++;
 469 out_set:
 470         set_bit(CACHE_HASHED, &item->h.flags);
 471         item->h.next = *head;
 472         *head = &item->h;
 473         write_unlock(&rsc_cache.hash_lock);
 474         cache_fresh(&rsc_cache, &item->h, item->h.expiry_time);
 475         cache_get(&item->h);
 476         RETURN(item);
 477 out_nada:
 478         tmp = NULL;
 479 out_noset:
 480         read_unlock(&rsc_cache.hash_lock);
 481         RETURN(tmp);
 482 }
 483
 484 static int rsc_parse(struct cache_detail *cd,
 485                      char *mesg, int mlen)
 486 {
 487         /* contexthandle expiry [ uid gid N <n gids> mechname
 488          * ...mechdata... ] */
 489         char *buf = mesg;
 490         int len, rv, tmp_int;
 491         struct rsc *rsci, *res = NULL;
 492         time_t expiry;
 493         int status = -EINVAL;
 494
 495         OBD_ALLOC(rsci, sizeof(*rsci));
 496         if (!rsci) {
 497                 CERROR("fail to alloc rsci\n");
 498                 return -ENOMEM;
 499         }
 500         cache_init(&rsci->h);
 501
 502         /* context handle */
 503         len = qword_get(&mesg, buf, mlen);
 504         if (len < 0) goto out;
 505         status = -ENOMEM;
 506         if (rawobj_alloc(&rsci->handle, buf, len))
 507                 goto out;
 508
 509         /* expiry */
 510         expiry = get_expiry(&mesg);
 511         status = -EINVAL;
 512         if (expiry == 0)
 513                 goto out;
 514
 515         /* remote flag */
 516         rv = get_int(&mesg, &tmp_int);
 517         if (rv) {
 518                 CERROR("fail to get remote flag\n");
 519                 goto out;
 520         }
 521         rsci->remote_realm = (tmp_int != 0);
 522
 523         /* mds user flag */
 524         rv = get_int(&mesg, &tmp_int);
 525         if (rv) {
 526                 CERROR("fail to get mds user flag\n");
 527                 goto out;
 528         }
 529         rsci->auth_usr_mds = (tmp_int != 0);
 530
 531         /* oss user flag */
 532         rv = get_int(&mesg, &tmp_int);
 533         if (rv) {
 534                 CERROR("fail to get oss user flag\n");
 535                 goto out;
 536         }
 537         rsci->auth_usr_oss = (tmp_int != 0);
 538
 539         /* mapped uid */
 540         rv = get_int(&mesg, (int *)&rsci->mapped_uid);
 541         if (rv) {
 542                 CERROR("fail to get mapped uid\n");
 543                 goto out;
 544         }
 545
 546         /* uid, or NEGATIVE */
 547         rv = get_int(&mesg, (int *)&rsci->cred.vc_uid);
 548         if (rv == -EINVAL)
 549                 goto out;
 550         if (rv == -ENOENT) {
 551                 CERROR("NOENT? set rsc entry negative\n");
 552                 set_bit(CACHE_NEGATIVE, &rsci->h.flags);
 553         } else {
 554                 struct gss_api_mech *gm;
 555                 rawobj_t tmp_buf;
 556                 __u64 ctx_expiry;
 557
 558                 /* gid */
 559                 if (get_int(&mesg, (int *)&rsci->cred.vc_gid))
 560                         goto out;
 561
 562                 /* mech name */
 563                 len = qword_get(&mesg, buf, mlen);
 564                 if (len < 0)
 565                         goto out;
 566                 gm = kgss_name_to_mech(buf);
 567                 status = -EOPNOTSUPP;
 568                 if (!gm)
 569                         goto out;
 570
 571                 status = -EINVAL;
 572                 /* mech-specific data: */
 573                 len = qword_get(&mesg, buf, mlen);
 574                 if (len < 0) {
 575                         kgss_mech_put(gm);
 576                         goto out;
 577                 }
 578                 tmp_buf.len = len;
 579                 tmp_buf.data = (unsigned char *)buf;
 580                 if (kgss_import_sec_context(&tmp_buf, gm, &rsci->mechctx)) {
 581                         kgss_mech_put(gm);
 582                         goto out;
 583                 }
 584
 585                 /* currently the expiry time passed down from user-space
 586                  * is invalid, here we retrive it from mech.
 587                  */
 588                 if (kgss_inquire_context(rsci->mechctx, &ctx_expiry)) {
 589                         CERROR("unable to get expire time, drop it\n");
 590                         set_bit(CACHE_NEGATIVE, &rsci->h.flags);
 591                         kgss_mech_put(gm);
 592                         goto out;
 593                 }
 594                 expiry = (time_t) gss_roundup_expire_time(ctx_expiry);
 595
 596                 kgss_mech_put(gm);
 597         }
 598         rsci->h.expiry_time = expiry;
 599         spin_lock_init(&rsci->seqdata.sd_lock);
 600         res = rsc_lookup(rsci, 1);
 601         rsc_put(&res->h, &rsc_cache);
 602         status = 0;
 603 out:
 604         if (rsci)
 605                 rsc_put(&rsci->h, &rsc_cache);
 606         return status;
 607 }
 608
 609 /*
 610  * flush all entries with @uid. @uid == -1 will match all.
 611  * we only know the uid, maybe netid/nid in the future, in all cases
 612  * we must search the whole cache
 613  */
 614 static void rsc_flush(uid_t uid)
 615 {
 616         struct cache_head **ch;
 617         struct rsc *rscp;
 618         int n;
 619         ENTRY;
 620
 621         if (uid == -1)
 622                 CWARN("flush all gss contexts\n");
 623
 624         write_lock(&rsc_cache.hash_lock);
 625         for (n = 0; n < RSC_HASHMAX; n++) {
 626                 for (ch = &rsc_cache.hash_table[n]; *ch;) {
 627                         rscp = container_of(*ch, struct rsc, h);
 628
 629                         if (uid != -1 && rscp->cred.vc_uid != uid) {
 630                                 ch = &((*ch)->next);
 631                                 continue;
 632                         }
 633
 634                         /* it seems simply set NEGATIVE doesn't work */
 635                         *ch = (*ch)->next;
 636                         rscp->h.next = NULL;
 637                         cache_get(&rscp->h);
 638                         set_bit(CACHE_NEGATIVE, &rscp->h.flags);
 639                         clear_bit(CACHE_HASHED, &rscp->h.flags);
 640                         if (uid != -1)
 641                                 CWARN("flush rsc %p(%u) for uid %u\n", rscp,
 642                                       *((__u32 *) rscp->handle.data),
 643                                       rscp->cred.vc_uid);
 644                         rsc_put(&rscp->h, &rsc_cache);
 645                         rsc_cache.entries--;
 646                 }
 647         }
 648         write_unlock(&rsc_cache.hash_lock);
 649         EXIT;
 650 }
 651
 652 static struct cache_detail rsc_cache = {
 653         .hash_size      = RSC_HASHMAX,
 654         .hash_table     = rsc_table,
 655         .name           = "auth.ptlrpcs.context",
 656         .cache_put      = rsc_put,
 657         .cache_parse    = rsc_parse,
 658 };
 659
 660 static struct rsc *
 661 gss_svc_searchbyctx(rawobj_t *handle)
 662 {
 663         struct rsc rsci;
 664         struct rsc *found;
 665
 666         rsci.handle = *handle;
 667         found = rsc_lookup(&rsci, 0);
 668         if (!found)
 669                 return NULL;
 670
 671         if (cache_check(&rsc_cache, &found->h, NULL))
 672                 return NULL;
 673
 674         return found;
 675 }
 676
 677 /* FIXME
 678  * again hacking: only try to give the svcgssd a chance to handle
 679  * upcalls.
 680  */
 681 struct cache_deferred_req* my_defer(struct cache_req *req)
 682 {
 683         yield();
 684         return NULL;
 685 }
 686 static struct cache_req my_chandle = {my_defer};
 687
 688 /* Implements sequence number algorithm as specified in RFC 2203. */
 689 static int
 690 gss_check_seq_num(struct gss_svc_seq_data *sd, __u32 seq_num)
 691 {
 692         int rc = 0;
 693
 694         spin_lock(&sd->sd_lock);
 695         if (seq_num > sd->sd_max) {
 696                 if (seq_num >= sd->sd_max + GSS_SEQ_WIN) {
 697                         memset(sd->sd_win, 0, sizeof(sd->sd_win));
 698                         sd->sd_max = seq_num;
 699                 } else {
 700                         while(sd->sd_max < seq_num) {
 701                                 sd->sd_max++;
 702                                 __clear_bit(sd->sd_max % GSS_SEQ_WIN,
 703                                             sd->sd_win);
 704                         }
 705                 }
 706                 __set_bit(seq_num % GSS_SEQ_WIN, sd->sd_win);
 707                 goto exit;
 708         } else if (seq_num + GSS_SEQ_WIN <= sd->sd_max) {
 709                 CERROR("seq %u too low: max %u, win %d\n",
 710                         seq_num, sd->sd_max, GSS_SEQ_WIN);
 711                 rc = 1;
 712                 goto exit;
 713         }
 714
 715         if (__test_and_set_bit(seq_num % GSS_SEQ_WIN, sd->sd_win)) {
 716                 CERROR("seq %u is replay: max %u, win %d\n",
 717                         seq_num, sd->sd_max, GSS_SEQ_WIN);
 718                 rc = 1;
 719         }
 720 exit:
 721         spin_unlock(&sd->sd_lock);
 722         return rc;
 723 }
 724
 725 static int
 726 gss_svc_verify_request(struct ptlrpc_request *req,
 727                        struct rsc *rsci,
 728                        struct rpc_gss_wire_cred *gc,
 729                        __u32 *vp, __u32 vlen)
 730 {
 731         struct ptlrpcs_wire_hdr *sec_hdr;
 732         struct gss_ctx *ctx = rsci->mechctx;
 733         __u32 maj_stat;
 734         rawobj_t msg;
 735         rawobj_t mic;
 736         ENTRY;
 737
 738         sec_hdr = (struct ptlrpcs_wire_hdr *) req->rq_reqbuf;
 739
 740         req->rq_reqmsg = (struct lustre_msg *) (req->rq_reqbuf + sizeof(*sec_hdr));
 741         req->rq_reqlen = sec_hdr->msg_len;
 742
 743         msg.len = sec_hdr->msg_len;
 744         msg.data = (__u8 *)req->rq_reqmsg;
 745
 746         mic.len = le32_to_cpu(*vp++);
 747         mic.data = (unsigned char *)vp;
 748         vlen -= 4;
 749
 750         if (mic.len > vlen) {
 751                 CERROR("checksum len %d, while buffer len %d\n",
 752                         mic.len, vlen);
 753                 RETURN(GSS_S_CALL_BAD_STRUCTURE);
 754         }
 755
 756         if (mic.len > 256) {
 757                 CERROR("invalid mic len %d\n", mic.len);
 758                 RETURN(GSS_S_CALL_BAD_STRUCTURE);
 759         }
 760
 761         maj_stat = kgss_verify_mic(ctx, &msg, &mic, NULL);
 762         if (maj_stat != GSS_S_COMPLETE) {
 763                 CERROR("MIC verification error: major %x\n", maj_stat);
 764                 RETURN(maj_stat);
 765         }
 766
 767         if (gss_check_seq_num(&rsci->seqdata, gc->gc_seq)) {
 768                 CERROR("discard replayed request %p(o%u,x"LPU64",t"LPU64")\n",
 769                         req, req->rq_reqmsg->opc, req->rq_xid,
 770                         req->rq_reqmsg->transno);
 771                 RETURN(GSS_S_DUPLICATE_TOKEN);
 772         }
 773
 774         RETURN(GSS_S_COMPLETE);
 775 }
 776
 777 static int
 778 gss_svc_unseal_request(struct ptlrpc_request *req,
 779                        struct rsc *rsci,
 780                        struct rpc_gss_wire_cred *gc,
 781                        __u32 *vp, __u32 vlen)
 782 {
 783         struct ptlrpcs_wire_hdr *sec_hdr;
 784         struct gss_ctx *ctx = rsci->mechctx;
 785         rawobj_t cipher_text, plain_text;
 786         __u32 major;
 787         ENTRY;
 788
 789         sec_hdr = (struct ptlrpcs_wire_hdr *) req->rq_reqbuf;
 790
 791         if (vlen < 4) {
 792                 CERROR("vlen only %u\n", vlen);
 793                 RETURN(GSS_S_CALL_BAD_STRUCTURE);
 794         }
 795
 796         cipher_text.len = le32_to_cpu(*vp++);
 797         cipher_text.data = (__u8 *) vp;
 798         vlen -= 4;
 799
 800         if (cipher_text.len > vlen) {
 801                 CERROR("cipher claimed %u while buf only %u\n",
 802                         cipher_text.len, vlen);
 803                 RETURN(GSS_S_CALL_BAD_STRUCTURE);
 804         }
 805
 806         plain_text = cipher_text;
 807
 808         major = kgss_unwrap(ctx, GSS_C_QOP_DEFAULT, &cipher_text, &plain_text);
 809         if (major) {
 810                 CERROR("unwrap error 0x%x\n", major);
 811                 RETURN(major);
 812         }
 813
 814         if (gss_check_seq_num(&rsci->seqdata, gc->gc_seq)) {
 815                 CERROR("discard replayed request %p(o%u,x"LPU64",t"LPU64")\n",
 816                         req, req->rq_reqmsg->opc, req->rq_xid,
 817                         req->rq_reqmsg->transno);
 818                 RETURN(GSS_S_DUPLICATE_TOKEN);
 819         }
 820
 821         req->rq_reqmsg = (struct lustre_msg *) (vp);
 822         req->rq_reqlen = plain_text.len;
 823
 824         CDEBUG(D_SEC, "msg len %d\n", req->rq_reqlen);
 825
 826         RETURN(GSS_S_COMPLETE);
 827 }
 828
 829 static int
 830 gss_pack_err_notify(struct ptlrpc_request *req,
 831                     __u32 major, __u32 minor)
 832 {
 833         struct gss_svc_data *svcdata = req->rq_svcsec_data;
 834         __u32 reslen, *resp, *reslenp;
 835         char  nidstr[PTL_NALFMT_SIZE];
 836         const __u32 secdata_len = 7 * 4;
 837         int rc;
 838         ENTRY;
 839
 840         OBD_FAIL_RETURN(OBD_FAIL_SVCGSS_ERR_NOTIFY|OBD_FAIL_ONCE, -EINVAL);
 841
 842         LASSERT(svcdata);
 843         svcdata->is_err_notify = 1;
 844         svcdata->reserve_len = 7 * 4;
 845
 846         rc = lustre_pack_reply(req, 0, NULL, NULL);
 847         if (rc) {
 848                 CERROR("could not pack reply, err %d\n", rc);
 849                 RETURN(rc);
 850         }
 851
 852         LASSERT(req->rq_reply_state);
 853         LASSERT(req->rq_reply_state->rs_repbuf);
 854         LASSERT(req->rq_reply_state->rs_repbuf_len >= secdata_len);
 855         resp = (__u32 *) req->rq_reply_state->rs_repbuf;
 856
 857         /* header */
 858         *resp++ = cpu_to_le32(PTLRPCS_FLVR_GSS_NONE);
 859         *resp++ = cpu_to_le32(PTLRPCS_SVC_NONE);
 860         *resp++ = cpu_to_le32(req->rq_replen);
 861         reslenp = resp++;
 862
 863         /* skip lustre msg */
 864         resp += req->rq_replen / 4;
 865         reslen = svcdata->reserve_len;
 866
 867         /* gss replay:
 868          * version, subflavor, notify, major, minor,
 869          * obj1(fake), obj2(fake)
 870          */
 871         *resp++ = cpu_to_le32(PTLRPC_SEC_GSS_VERSION);
 872         *resp++ = cpu_to_le32(PTLRPCS_FLVR_KRB5I);
 873         *resp++ = cpu_to_le32(PTLRPCS_GSS_PROC_ERR);
 874         *resp++ = cpu_to_le32(major);
 875         *resp++ = cpu_to_le32(minor);
 876         *resp++ = 0;
 877         *resp++ = 0;
 878         reslen -= (4 * 4);
 879         /* the actual sec data length */
 880         *reslenp = cpu_to_le32(secdata_len);
 881
 882         req->rq_reply_state->rs_repdata_len += (secdata_len);
 883         CDEBUG(D_SEC, "prepare gss error notify(0x%x/0x%x) to %s\n",
 884                major, minor,
 885                portals_nid2str(req->rq_peer.peer_ni->pni_number,
 886                                req->rq_peer.peer_id.nid, nidstr));
 887         RETURN(0);
 888 }
 889
 890 static void dump_cache_head(struct cache_head *h)
 891 {
 892         CWARN("ref %d, fl %lx, n %p, t %ld, %ld\n",
 893               atomic_read(&h->refcnt), h->flags, h->next,
 894               h->expiry_time, h->last_refresh);
 895 }
 896 static void dump_rsi(struct rsi *rsi)
 897 {
 898         CWARN("dump rsi %p\n", rsi);
 899         dump_cache_head(&rsi->h);
 900         CWARN("%x,%x,%llx\n", rsi->naltype, rsi->netid, rsi->nid);
 901         CWARN("len %d, d %p\n", rsi->in_handle.len, rsi->in_handle.data);
 902         CWARN("len %d, d %p\n", rsi->in_token.len, rsi->in_token.data);
 903         CWARN("len %d, d %p\n", rsi->out_handle.len, rsi->out_handle.data);
 904         CWARN("len %d, d %p\n", rsi->out_token.len, rsi->out_token.data);
 905 }
 906
 907 static int
 908 gss_svcsec_handle_init(struct ptlrpc_request *req,
 909                        struct rpc_gss_wire_cred *gc,
 910                        __u32 *secdata, __u32 seclen,
 911                        enum ptlrpcs_error *res)
 912 {
 913         struct gss_svc_data *svcdata = req->rq_svcsec_data;
 914         struct rsc          *rsci;
 915         struct rsi          *rsikey, *rsip;
 916         rawobj_t             tmpobj;
 917         __u32 reslen,       *resp, *reslenp;
 918         char                 nidstr[PTL_NALFMT_SIZE];
 919         int                  rc;
 920         ENTRY;
 921
 922         LASSERT(svcdata);
 923
 924         CDEBUG(D_SEC, "processing gss init(%d) request from %s\n", gc->gc_proc,
 925                portals_nid2str(req->rq_peer.peer_ni->pni_number,
 926                                req->rq_peer.peer_id.nid, nidstr));
 927
 928         *res = PTLRPCS_BADCRED;
 929         OBD_FAIL_RETURN(OBD_FAIL_SVCGSS_INIT_REQ|OBD_FAIL_ONCE, SVC_DROP);
 930
 931         if (gc->gc_proc == RPC_GSS_PROC_INIT &&
 932             gc->gc_ctx.len != 0) {
 933                 CERROR("proc %d, ctx_len %d: not really init?\n",
 934                 gc->gc_proc == RPC_GSS_PROC_INIT, gc->gc_ctx.len);
 935                 RETURN(SVC_DROP);
 936         }
 937
 938         OBD_ALLOC(rsikey, sizeof(*rsikey));
 939         if (!rsikey) {
 940                 CERROR("out of memory\n");
 941                 RETURN(SVC_DROP);
 942         }
 943         cache_init(&rsikey->h);
 944
 945         /* obtain lustre svc type */
 946         if (seclen < 4) {
 947                 CERROR("sec size %d too small\n", seclen);
 948                 GOTO(out_rsikey, rc = SVC_DROP);
 949         }
 950         rsikey->lustre_svc = le32_to_cpu(*secdata++);
 951         seclen -= 4;
 952
 953         /* duplicate context handle. currently always 0 */
 954         if (rawobj_dup(&rsikey->in_handle, &gc->gc_ctx)) {
 955                 CERROR("fail to dup context handle\n");
 956                 GOTO(out_rsikey, rc = SVC_DROP);
 957         }
 958
 959         /* extract token */
 960         *res = PTLRPCS_BADVERF;
 961         if (rawobj_extract(&tmpobj, &secdata, &seclen)) {
 962                 CERROR("can't extract token\n");
 963                 GOTO(out_rsikey, rc = SVC_DROP);
 964         }
 965         if (rawobj_dup(&rsikey->in_token, &tmpobj)) {
 966                 CERROR("can't duplicate token\n");
 967                 GOTO(out_rsikey, rc = SVC_DROP);
 968         }
 969
 970         rsikey->naltype = (__u32) req->rq_peer.peer_ni->pni_number;
 971         rsikey->netid = 0;
 972         rsikey->nid = (__u64) req->rq_peer.peer_id.nid;
 973
 974         rsip = gssd_upcall(rsikey, &my_chandle);
 975         if (!rsip) {
 976                 CERROR("error in gssd_upcall.\n");
 977
 978                 rc = SVC_COMPLETE;
 979                 if (gss_pack_err_notify(req, GSS_S_FAILURE, 0))
 980                         rc = SVC_DROP;
 981
 982                 GOTO(out_rsikey, rc);
 983         }
 984
 985         rsci = gss_svc_searchbyctx(&rsip->out_handle);
 986         if (!rsci) {
 987                 CERROR("rsci still not mature yet?\n");
 988
 989                 rc = SVC_COMPLETE;
 990                 if (gss_pack_err_notify(req, GSS_S_FAILURE, 0))
 991                         rc = SVC_DROP;
 992
 993                 GOTO(out_rsip, rc);
 994         }
 995         CDEBUG(D_SEC, "svcsec create gss context %p(%u@%s)\n",
 996                rsci, rsci->cred.vc_uid,
 997                portals_nid2str(req->rq_peer.peer_ni->pni_number,
 998                                req->rq_peer.peer_id.nid, nidstr));
 999
1000         svcdata->is_init = 1;
1001         svcdata->reserve_len = 7 * 4 +
1002                 size_round4(rsip->out_handle.len) +
1003                 size_round4(rsip->out_token.len);
1004
1005         rc = lustre_pack_reply(req, 0, NULL, NULL);
1006         if (rc) {
1007                 CERROR("failed to pack reply, rc = %d\n", rc);
1008                 set_bit(CACHE_NEGATIVE, &rsci->h.flags);
1009                 GOTO(out, rc = SVC_DROP);
1010         }
1011
1012         /* header */
1013         resp = (__u32 *) req->rq_reply_state->rs_repbuf;
1014         *resp++ = cpu_to_le32(PTLRPCS_FLVR_GSS_NONE);
1015         *resp++ = cpu_to_le32(PTLRPCS_SVC_NONE);
1016         *resp++ = cpu_to_le32(req->rq_replen);
1017         reslenp = resp++;
1018
1019         resp += req->rq_replen / 4;
1020         reslen = svcdata->reserve_len;
1021
1022         /* gss reply: (conform to err notify format)
1023          * x, x, seq, major, minor, handle, token
1024          */
1025         *resp++ = 0;
1026         *resp++ = 0;
1027         *resp++ = cpu_to_le32(GSS_SEQ_WIN);
1028         *resp++ = cpu_to_le32(rsip->major_status);
1029         *resp++ = cpu_to_le32(rsip->minor_status);
1030         reslen -= (5 * 4);
1031         if (rawobj_serialize(&rsip->out_handle,
1032                              &resp, &reslen)) {
1033                 dump_rsi(rsip);
1034                 dump_rsi(rsikey);
1035                 LBUG();
1036         }
1037         if (rawobj_serialize(&rsip->out_token,
1038                              &resp, &reslen)) {
1039                 dump_rsi(rsip);
1040                 dump_rsi(rsikey);
1041                 LBUG();
1042         }
1043         /* the actual sec data length */
1044         *reslenp = cpu_to_le32(svcdata->reserve_len - reslen);
1045
1046         req->rq_reply_state->rs_repdata_len += le32_to_cpu(*reslenp);
1047         CDEBUG(D_SEC, "req %p: msgsize %d, authsize %d, "
1048                "total size %d\n", req, req->rq_replen,
1049                le32_to_cpu(*reslenp),
1050                req->rq_reply_state->rs_repdata_len);
1051
1052         *res = PTLRPCS_OK;
1053
1054         req->rq_remote_realm = rsci->remote_realm;
1055         req->rq_auth_usr_mds = rsci->auth_usr_mds;
1056         req->rq_auth_usr_oss = rsci->auth_usr_oss;
1057         req->rq_auth_uid = rsci->cred.vc_uid;
1058         req->rq_mapped_uid = rsci->mapped_uid;
1059
1060         if (req->rq_auth_usr_mds) {
1061                 CWARN("usr from %s authenticated as mds svc cred\n",
1062                 portals_nid2str(req->rq_peer.peer_ni->pni_number,
1063                                 req->rq_peer.peer_id.nid, nidstr));
1064         }
1065         if (req->rq_auth_usr_oss) {
1066                 CWARN("usr from %s authenticated as oss svc cred\n",
1067                 portals_nid2str(req->rq_peer.peer_ni->pni_number,
1068                                 req->rq_peer.peer_id.nid, nidstr));
1069         }
1070
1071         /* This is simplified since right now we doesn't support
1072          * INIT_CONTINUE yet.
1073          */
1074         if (gc->gc_proc == RPC_GSS_PROC_INIT) {
1075                 struct ptlrpcs_wire_hdr *hdr;
1076
1077                 hdr = buf_to_sec_hdr(req->rq_reqbuf);
1078                 req->rq_reqmsg = buf_to_lustre_msg(req->rq_reqbuf);
1079                 req->rq_reqlen = hdr->msg_len;
1080
1081                 rc = SVC_LOGIN;
1082         } else
1083                 rc = SVC_COMPLETE;
1084
1085 out:
1086         rsc_put(&rsci->h, &rsc_cache);
1087 out_rsip:
1088         rsi_put(&rsip->h, &rsi_cache);
1089 out_rsikey:
1090         rsi_put(&rsikey->h, &rsi_cache);
1091
1092         RETURN(rc);
1093 }
1094
1095 static int
1096 gss_svcsec_handle_data(struct ptlrpc_request *req,
1097                        struct rpc_gss_wire_cred *gc,
1098                        __u32 *secdata, __u32 seclen,
1099                        enum ptlrpcs_error *res)
1100 {
1101         struct rsc          *rsci;
1102         char                 nidstr[PTL_NALFMT_SIZE];
1103         __u32                major;
1104         int                  rc;
1105         ENTRY;
1106
1107         *res = PTLRPCS_GSS_CREDPROBLEM;
1108
1109         rsci = gss_svc_searchbyctx(&gc->gc_ctx);
1110         if (!rsci) {
1111                 CWARN("Invalid gss context handle from %s\n",
1112                        portals_nid2str(req->rq_peer.peer_ni->pni_number,
1113                                        req->rq_peer.peer_id.nid, nidstr));
1114                 major = GSS_S_NO_CONTEXT;
1115                 goto notify_err;
1116         }
1117
1118         switch (gc->gc_svc) {
1119         case PTLRPCS_GSS_SVC_INTEGRITY:
1120                 major = gss_svc_verify_request(req, rsci, gc, secdata, seclen);
1121                 if (major == GSS_S_COMPLETE)
1122                         break;
1123
1124                 CWARN("fail in verify: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         case PTLRPCS_GSS_SVC_PRIVACY:
1129                 major = gss_svc_unseal_request(req, rsci, gc, secdata, seclen);
1130                 if (major == GSS_S_COMPLETE)
1131                         break;
1132
1133                 CWARN("fail in decrypt:0x%x: ctx %p@%s\n", major, rsci,
1134                        portals_nid2str(req->rq_peer.peer_ni->pni_number,
1135                                        req->rq_peer.peer_id.nid, nidstr));
1136                 goto notify_err;
1137         default:
1138                 CERROR("unsupported gss service %d\n", gc->gc_svc);
1139                 GOTO(out, rc = SVC_DROP);
1140         }
1141
1142         req->rq_remote_realm = rsci->remote_realm;
1143         req->rq_auth_usr_mds = rsci->auth_usr_mds;
1144         req->rq_auth_usr_oss = rsci->auth_usr_oss;
1145         req->rq_auth_uid = rsci->cred.vc_uid;
1146         req->rq_mapped_uid = rsci->mapped_uid;
1147
1148         *res = PTLRPCS_OK;
1149         GOTO(out, rc = SVC_OK);
1150
1151 notify_err:
1152         if (gss_pack_err_notify(req, major, 0))
1153                 rc = SVC_DROP;
1154         else
1155                 rc = SVC_COMPLETE;
1156 out:
1157         if (rsci)
1158                 rsc_put(&rsci->h, &rsc_cache);
1159         RETURN(rc);
1160 }
1161
1162 static int
1163 gss_svcsec_handle_destroy(struct ptlrpc_request *req,
1164                           struct rpc_gss_wire_cred *gc,
1165                           __u32 *secdata, __u32 seclen,
1166                           enum ptlrpcs_error *res)
1167 {
1168         struct gss_svc_data *svcdata = req->rq_svcsec_data;
1169         struct rsc          *rsci;
1170         char                 nidstr[PTL_NALFMT_SIZE];
1171         int                  rc;
1172         ENTRY;
1173
1174         LASSERT(svcdata);
1175         *res = PTLRPCS_GSS_CREDPROBLEM;
1176
1177         rsci = gss_svc_searchbyctx(&gc->gc_ctx);
1178         if (!rsci) {
1179                 CWARN("invalid gss context handle for destroy.\n");
1180                 RETURN(SVC_DROP);
1181         }
1182
1183         if (gc->gc_svc != PTLRPCS_GSS_SVC_INTEGRITY) {
1184                 CERROR("service %d is not supported in destroy.\n",
1185                         gc->gc_svc);
1186                 GOTO(out, rc = SVC_DROP);
1187         }
1188
1189         *res = gss_svc_verify_request(req, rsci, gc, secdata, seclen);
1190         if (*res)
1191                 GOTO(out, rc = SVC_DROP);
1192
1193         /* compose reply, which is actually nothing */
1194         svcdata->is_fini = 1;
1195         if (lustre_pack_reply(req, 0, NULL, NULL))
1196                 GOTO(out, rc = SVC_DROP);
1197
1198         CDEBUG(D_SEC, "svcsec destroy gss context %p(%u@%s)\n",
1199                rsci, rsci->cred.vc_uid,
1200                portals_nid2str(req->rq_peer.peer_ni->pni_number,
1201                                req->rq_peer.peer_id.nid, nidstr));
1202
1203         set_bit(CACHE_NEGATIVE, &rsci->h.flags);
1204         *res = PTLRPCS_OK;
1205         rc = SVC_LOGOUT;
1206 out:
1207         rsc_put(&rsci->h, &rsc_cache);
1208         RETURN(rc);
1209 }
1210
1211 /*
1212  * let incomming request go through security check:
1213  *  o context establishment: invoke user space helper
1214  *  o data exchange: verify/decrypt
1215  *  o context destruction: mark context invalid
1216  *
1217  * in most cases, error will result to drop the packet silently.
1218  */
1219 static int
1220 gss_svcsec_accept(struct ptlrpc_request *req, enum ptlrpcs_error *res)
1221 {
1222         struct gss_svc_data *svcdata;
1223         struct rpc_gss_wire_cred *gc;
1224         struct ptlrpcs_wire_hdr *sec_hdr;
1225         __u32 subflavor, seclen, *secdata, version;
1226         int rc;
1227         ENTRY;
1228
1229         CDEBUG(D_SEC, "request %p\n", req);
1230         LASSERT(req->rq_reqbuf);
1231         LASSERT(req->rq_reqbuf_len);
1232
1233         *res = PTLRPCS_BADCRED;
1234
1235         sec_hdr = buf_to_sec_hdr(req->rq_reqbuf);
1236         LASSERT(SEC_FLAVOR_MAJOR(sec_hdr->flavor) == PTLRPCS_FLVR_MAJOR_GSS);
1237
1238         seclen = req->rq_reqbuf_len - sizeof(*sec_hdr) - sec_hdr->msg_len;
1239         secdata = (__u32 *) buf_to_sec_data(req->rq_reqbuf);
1240
1241         if (sec_hdr->sec_len > seclen) {
1242                 CERROR("seclen %d, while max buf %d\n",
1243                         sec_hdr->sec_len, seclen);
1244                 RETURN(SVC_DROP);
1245         }
1246
1247         if (seclen < 6 * 4) {
1248                 CERROR("sec size %d too small\n", seclen);
1249                 RETURN(SVC_DROP);
1250         }
1251
1252         LASSERT(!req->rq_svcsec_data);
1253         OBD_ALLOC(svcdata, sizeof(*svcdata));
1254         if (!svcdata) {
1255                 CERROR("fail to alloc svcdata\n");
1256                 RETURN(SVC_DROP);
1257         }
1258         req->rq_svcsec_data = svcdata;
1259         gc = &svcdata->clcred;
1260
1261         /* Now secdata/seclen is what we want to parse
1262          */
1263         version = le32_to_cpu(*secdata++);      /* version */
1264         subflavor = le32_to_cpu(*secdata++);    /* subflavor */
1265         gc->gc_proc = le32_to_cpu(*secdata++);  /* proc */
1266         gc->gc_seq = le32_to_cpu(*secdata++);   /* seq */
1267         gc->gc_svc = le32_to_cpu(*secdata++);   /* service */
1268         seclen -= 5 * 4;
1269
1270         CDEBUG(D_SEC, "wire gss_hdr: %u/%u/%u/%u/%u\n",
1271                version, subflavor, gc->gc_proc,
1272                gc->gc_seq, gc->gc_svc);
1273
1274         if (version != PTLRPC_SEC_GSS_VERSION) {
1275                 CERROR("gss version mismatch: %d - %d\n",
1276                         version, PTLRPC_SEC_GSS_VERSION);
1277                 GOTO(err_free, rc = SVC_DROP);
1278         }
1279
1280         /* We _must_ alloc new storage for gc_ctx. In case of recovery
1281          * request will be saved to delayed handling, at that time the
1282          * incoming buffer might have already been released.
1283          */
1284         if (rawobj_extract_alloc(&gc->gc_ctx, &secdata, &seclen)) {
1285                 CERROR("fail to obtain gss context handle\n");
1286                 GOTO(err_free, rc = SVC_DROP);
1287         }
1288
1289         *res = PTLRPCS_BADVERF;
1290         switch(gc->gc_proc) {
1291         case RPC_GSS_PROC_INIT:
1292         case RPC_GSS_PROC_CONTINUE_INIT:
1293                 rc = gss_svcsec_handle_init(req, gc, secdata, seclen, res);
1294                 break;
1295         case RPC_GSS_PROC_DATA:
1296                 rc = gss_svcsec_handle_data(req, gc, secdata, seclen, res);
1297                 break;
1298         case RPC_GSS_PROC_DESTROY:
1299                 rc = gss_svcsec_handle_destroy(req, gc, secdata, seclen, res);
1300                 break;
1301         default:
1302                 rc = SVC_DROP;
1303                 LBUG();
1304         }
1305
1306 err_free:
1307         if (rc == SVC_DROP && req->rq_svcsec_data) {
1308                 OBD_FREE(req->rq_svcsec_data, sizeof(struct gss_svc_data));
1309                 req->rq_svcsec_data = NULL;
1310         }
1311
1312         RETURN(rc);
1313 }
1314
1315 static int
1316 gss_svcsec_authorize(struct ptlrpc_request *req)
1317 {
1318         struct ptlrpc_reply_state *rs = req->rq_reply_state;
1319         struct gss_svc_data *gsd = (struct gss_svc_data *)req->rq_svcsec_data;
1320         struct rpc_gss_wire_cred  *gc = &gsd->clcred;
1321         struct rsc                *rscp;
1322         struct ptlrpcs_wire_hdr   *sec_hdr;
1323         rawobj_buf_t               msg_buf;
1324         rawobj_t                   cipher_buf;
1325         __u32                     *vp, *vpsave, major, vlen, seclen;
1326         rawobj_t                   lmsg, mic;
1327         int                        ret;
1328         ENTRY;
1329
1330         LASSERT(rs);
1331         LASSERT(rs->rs_repbuf);
1332         LASSERT(gsd);
1333
1334         if (gsd->is_init || gsd->is_init_continue ||
1335             gsd->is_err_notify || gsd->is_fini) {
1336                 /* nothing to do in these cases */
1337                 CDEBUG(D_SEC, "req %p: init/fini/err\n", req);
1338                 RETURN(0);
1339         }
1340
1341         if (gc->gc_proc != RPC_GSS_PROC_DATA) {
1342                 CERROR("proc %d not support\n", gc->gc_proc);
1343                 RETURN(-EINVAL);
1344         }
1345
1346         rscp = gss_svc_searchbyctx(&gc->gc_ctx);
1347         if (!rscp) {
1348                 CERROR("ctx %u disapeared under us\n",
1349                        *((__u32 *) gc->gc_ctx.data));
1350                 RETURN(-EINVAL);
1351         }
1352
1353         sec_hdr = (struct ptlrpcs_wire_hdr *) rs->rs_repbuf;
1354         switch (gc->gc_svc) {
1355         case  PTLRPCS_GSS_SVC_INTEGRITY:
1356                 /* prepare various pointers */
1357                 lmsg.len = req->rq_replen;
1358                 lmsg.data = (__u8 *) (rs->rs_repbuf + sizeof(*sec_hdr));
1359                 vp = (__u32 *) (lmsg.data + lmsg.len);
1360                 vlen = rs->rs_repbuf_len - sizeof(*sec_hdr) - lmsg.len;
1361                 seclen = vlen;
1362
1363                 sec_hdr->flavor = cpu_to_le32(PTLRPCS_FLVR_GSS_AUTH);
1364                 sec_hdr->msg_len = cpu_to_le32(req->rq_replen);
1365
1366                 /* standard gss hdr */
1367                 LASSERT(vlen >= 7 * 4);
1368                 *vp++ = cpu_to_le32(PTLRPC_SEC_GSS_VERSION);
1369                 *vp++ = cpu_to_le32(PTLRPCS_FLVR_KRB5I);
1370                 *vp++ = cpu_to_le32(RPC_GSS_PROC_DATA);
1371                 *vp++ = cpu_to_le32(gc->gc_seq);
1372                 *vp++ = cpu_to_le32(PTLRPCS_GSS_SVC_INTEGRITY);
1373                 *vp++ = 0;      /* fake ctx handle */
1374                 vpsave = vp++;  /* reserve size */
1375                 vlen -= 7 * 4;
1376
1377                 mic.len = vlen;
1378                 mic.data = (unsigned char *)vp;
1379
1380                 major = kgss_get_mic(rscp->mechctx, 0, &lmsg, &mic);
1381                 if (major) {
1382                         CERROR("fail to get MIC: 0x%x\n", major);
1383                         GOTO(out, ret = -EINVAL);
1384                 }
1385                 *vpsave = cpu_to_le32(mic.len);
1386                 seclen = seclen - vlen + mic.len;
1387                 sec_hdr->sec_len = cpu_to_le32(seclen);
1388                 rs->rs_repdata_len += size_round(seclen);
1389                 break;
1390         case  PTLRPCS_GSS_SVC_PRIVACY:
1391                 vp = (__u32 *) (rs->rs_repbuf + sizeof(*sec_hdr));
1392                 vlen = rs->rs_repbuf_len - sizeof(*sec_hdr);
1393                 seclen = vlen;
1394
1395                 sec_hdr->flavor = cpu_to_le32(PTLRPCS_FLVR_GSS_PRIV);
1396                 sec_hdr->msg_len = cpu_to_le32(0);
1397
1398                 /* standard gss hdr */
1399                 LASSERT(vlen >= 7 * 4);
1400                 *vp++ = cpu_to_le32(PTLRPC_SEC_GSS_VERSION);
1401                 *vp++ = cpu_to_le32(PTLRPCS_FLVR_KRB5I);
1402                 *vp++ = cpu_to_le32(RPC_GSS_PROC_DATA);
1403                 *vp++ = cpu_to_le32(gc->gc_seq);
1404                 *vp++ = cpu_to_le32(PTLRPCS_GSS_SVC_PRIVACY);
1405                 *vp++ = 0;      /* fake ctx handle */
1406                 vpsave = vp++;  /* reserve size */
1407                 vlen -= 7 * 4;
1408
1409                 msg_buf.buf = (__u8 *) rs->rs_msg - GSS_PRIVBUF_PREFIX_LEN;
1410                 msg_buf.buflen = req->rq_replen + GSS_PRIVBUF_PREFIX_LEN +
1411                                  GSS_PRIVBUF_SUFFIX_LEN;
1412                 msg_buf.dataoff = GSS_PRIVBUF_PREFIX_LEN;
1413                 msg_buf.datalen = req->rq_replen;
1414
1415                 cipher_buf.data = (__u8 *) vp;
1416                 cipher_buf.len = vlen;
1417
1418                 major = kgss_wrap(rscp->mechctx, GSS_C_QOP_DEFAULT,
1419                                 &msg_buf, &cipher_buf);
1420                 if (major) {
1421                         CERROR("failed to wrap: 0x%x\n", major);
1422                         GOTO(out, ret = -EINVAL);
1423                 }
1424
1425                 *vpsave = cpu_to_le32(cipher_buf.len);
1426                 seclen = seclen - vlen + cipher_buf.len;
1427                 sec_hdr->sec_len = cpu_to_le32(seclen);
1428                 rs->rs_repdata_len += size_round(seclen);
1429                 break;
1430         default:
1431                 CERROR("Unknown service %d\n", gc->gc_svc);
1432                 GOTO(out, ret = -EINVAL);
1433         }
1434         ret = 0;
1435 out:
1436         rsc_put(&rscp->h, &rsc_cache);
1437
1438         RETURN(ret);
1439 }
1440
1441 static
1442 void gss_svcsec_cleanup_req(struct ptlrpc_svcsec *svcsec,
1443                             struct ptlrpc_request *req)
1444 {
1445         struct gss_svc_data *gsd = (struct gss_svc_data *) req->rq_svcsec_data;
1446
1447         if (!gsd) {
1448                 CDEBUG(D_SEC, "no svc_data present. do nothing\n");
1449                 return;
1450         }
1451
1452         /* gc_ctx is allocated, see gss_svcsec_accept() */
1453         rawobj_free(&gsd->clcred.gc_ctx);
1454
1455         OBD_FREE(gsd, sizeof(*gsd));
1456         req->rq_svcsec_data = NULL;
1457         return;
1458 }
1459
1460 static
1461 int gss_svcsec_est_payload(struct ptlrpc_svcsec *svcsec,
1462                            struct ptlrpc_request *req,
1463                            int msgsize)
1464 {
1465         struct gss_svc_data *svcdata = req->rq_svcsec_data;
1466         ENTRY;
1467
1468         /* just return the pre-set reserve_len for init/fini/err cases.
1469          */
1470         LASSERT(svcdata);
1471         if (svcdata->is_init) {
1472                 CDEBUG(D_SEC, "is_init, reserver size %d(%d)\n",
1473                        size_round(svcdata->reserve_len),
1474                        svcdata->reserve_len);
1475                 LASSERT(svcdata->reserve_len);
1476                 LASSERT(svcdata->reserve_len % 4 == 0);
1477                 RETURN(size_round(svcdata->reserve_len));
1478         } else if (svcdata->is_err_notify) {
1479                 CDEBUG(D_SEC, "is_err_notify, reserver size %d(%d)\n",
1480                        size_round(svcdata->reserve_len),
1481                        svcdata->reserve_len);
1482                 RETURN(size_round(svcdata->reserve_len));
1483         } else if (svcdata->is_fini) {
1484                 CDEBUG(D_SEC, "is_fini, reserver size 0\n");
1485                 RETURN(0);
1486         } else {
1487                 if (svcdata->clcred.gc_svc == PTLRPCS_GSS_SVC_NONE ||
1488                     svcdata->clcred.gc_svc == PTLRPCS_GSS_SVC_INTEGRITY)
1489                         RETURN(size_round(GSS_MAX_AUTH_PAYLOAD));
1490                 else if (svcdata->clcred.gc_svc == PTLRPCS_GSS_SVC_PRIVACY)
1491                         RETURN(size_round16(GSS_MAX_AUTH_PAYLOAD + msgsize +
1492                                             GSS_PRIVBUF_PREFIX_LEN +
1493                                             GSS_PRIVBUF_SUFFIX_LEN));
1494                 else {
1495                         CERROR("unknown gss svc %u\n", svcdata->clcred.gc_svc);
1496                         *((int *)0) = 0;
1497                         LBUG();
1498                 }
1499         }
1500         RETURN(0);
1501 }
1502
1503 int gss_svcsec_alloc_repbuf(struct ptlrpc_svcsec *svcsec,
1504                             struct ptlrpc_request *req,
1505                             int msgsize)
1506 {
1507         struct gss_svc_data *gsd = (struct gss_svc_data *) req->rq_svcsec_data;
1508         struct ptlrpc_reply_state *rs;
1509         int msg_payload, sec_payload;
1510         int privacy, rc;
1511         ENTRY;
1512
1513         /* determine the security type: none/auth or priv, we have
1514          * different pack scheme for them.
1515          * init/fini/err will always be treated as none/auth.
1516          */
1517         LASSERT(gsd);
1518         if (!gsd->is_init && !gsd->is_init_continue &&
1519             !gsd->is_fini && !gsd->is_err_notify &&
1520             gsd->clcred.gc_svc == PTLRPCS_GSS_SVC_PRIVACY)
1521                 privacy = 1;
1522         else
1523                 privacy = 0;
1524
1525         msg_payload = privacy ? 0 : msgsize;
1526         sec_payload = gss_svcsec_est_payload(svcsec, req, msgsize);
1527
1528         rc = svcsec_alloc_reply_state(req, msg_payload, sec_payload);
1529         if (rc)
1530                 RETURN(rc);
1531
1532         rs = req->rq_reply_state;
1533         LASSERT(rs);
1534         rs->rs_msg_len = msgsize;
1535
1536         if (privacy) {
1537                 /* we can choose to let msg simply point to the rear of the
1538                  * buffer, which lead to buffer overlap when doing encryption.
1539                  * usually it's ok and it indeed passed all existing tests.
1540                  * but not sure if there will be subtle problems in the future.
1541                  * so right now we choose to alloc another new buffer. we'll
1542                  * see how it works.
1543                  */
1544 #if 0
1545                 rs->rs_msg = (struct lustre_msg *)
1546                              (rs->rs_repbuf + rs->rs_repbuf_len -
1547                               msgsize - GSS_PRIVBUF_SUFFIX_LEN);
1548 #endif
1549                 char *msgbuf;
1550
1551                 msgsize += GSS_PRIVBUF_PREFIX_LEN + GSS_PRIVBUF_SUFFIX_LEN;
1552                 OBD_ALLOC(msgbuf, msgsize);
1553                 if (!msgbuf) {
1554                         CERROR("can't alloc %d\n", msgsize);
1555                         svcsec_free_reply_state(rs);
1556                         req->rq_reply_state = NULL;
1557                         RETURN(-ENOMEM);
1558                 }
1559                 rs->rs_msg = (struct lustre_msg *)
1560                                 (msgbuf + GSS_PRIVBUF_PREFIX_LEN);
1561         }
1562
1563         req->rq_repmsg = rs->rs_msg;
1564
1565         RETURN(0);
1566 }
1567
1568 static
1569 void gss_svcsec_free_repbuf(struct ptlrpc_svcsec *svcsec,
1570                             struct ptlrpc_reply_state *rs)
1571 {
1572         unsigned long p1 = (unsigned long) rs->rs_msg;
1573         unsigned long p2 = (unsigned long) rs->rs_buf;
1574
1575         LASSERT(rs->rs_buf);
1576         LASSERT(rs->rs_msg);
1577         LASSERT(rs->rs_msg_len);
1578
1579         if (p1 < p2 || p1 >= p2 + rs->rs_buf_len) {
1580                 char *start = (char*) rs->rs_msg - GSS_PRIVBUF_PREFIX_LEN;
1581                 int size = rs->rs_msg_len + GSS_PRIVBUF_PREFIX_LEN +
1582                            GSS_PRIVBUF_SUFFIX_LEN;
1583                 OBD_FREE(start, size);
1584         }
1585
1586         svcsec_free_reply_state(rs);
1587 }
1588
1589 struct ptlrpc_svcsec svcsec_gss = {
1590         .pss_owner              = THIS_MODULE,
1591         .pss_name               = "svcsec.gss",
1592         .pss_flavor             = PTLRPCS_FLVR_MAJOR_GSS,
1593         .accept                 = gss_svcsec_accept,
1594         .authorize              = gss_svcsec_authorize,
1595         .alloc_repbuf           = gss_svcsec_alloc_repbuf,
1596         .free_repbuf            = gss_svcsec_free_repbuf,
1597         .cleanup_req            = gss_svcsec_cleanup_req,
1598 };
1599
1600 /* XXX hacking */
1601 void lgss_svc_cache_purge_all(void)
1602 {
1603         cache_purge(&rsi_cache);
1604         cache_purge(&rsc_cache);
1605 }
1606 EXPORT_SYMBOL(lgss_svc_cache_purge_all);
1607
1608 void lgss_svc_cache_flush(__u32 uid)
1609 {
1610         rsc_flush(uid);
1611 }
1612 EXPORT_SYMBOL(lgss_svc_cache_flush);
1613
1614 int gss_svc_init(void)
1615 {
1616         int rc;
1617
1618         rc = svcsec_register(&svcsec_gss);
1619         if (!rc) {
1620                 cache_register(&rsc_cache);
1621                 cache_register(&rsi_cache);
1622         }
1623         return rc;
1624 }
1625
1626 void gss_svc_exit(void)
1627 {
1628         int rc;
1629
1630         /* XXX rsi didn't take module refcount. without really
1631          * cleanup it we can't simply go, later user-space operations
1632          * will certainly cause oops.
1633          * use space might slow or stuck on something, wait it for
1634          * a bit -- bad hack.
1635          */
1636         while ((rc = cache_unregister(&rsi_cache))) {
1637                 CERROR("unregister rsi cache: %d. Try again\n", rc);
1638                 schedule_timeout(2 * HZ);
1639                 cache_purge(&rsi_cache);
1640         }
1641
1642         if ((rc = cache_unregister(&rsc_cache)))
1643                 CERROR("unregister rsc cache: %d\n", rc);
1644         if ((rc = svcsec_unregister(&svcsec_gss)))
1645                 CERROR("unregister svcsec_gss: %d\n", rc);
1646 }