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) ((__u32) 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         if (uid == -1)
 621                 CWARN("flush all gss contexts\n");
 622
 623         write_lock(&rsc_cache.hash_lock);
 624         for (n = 0; n < RSC_HASHMAX; n++) {
 625                 for (ch = &rsc_cache.hash_table[n]; *ch;) {
 626                         rscp = container_of(*ch, struct rsc, h);
 627
 628                         if (uid != -1 && rscp->cred.vc_uid != uid) {
 629                                 ch = &((*ch)->next);
 630                                 continue;
 631                         }
 632
 633                         /* it seems simply set NEGATIVE doesn't work */
 634                         *ch = (*ch)->next;
 635                         rscp->h.next = NULL;
 636                         cache_get(&rscp->h);
 637                         set_bit(CACHE_NEGATIVE, &rscp->h.flags);
 638                         clear_bit(CACHE_HASHED, &rscp->h.flags);
 639                         if (uid != -1)
 640                                 CWARN("flush rsc %p(%u) for uid %u\n", rscp,
 641                                       *((__u32 *) rscp->handle.data),
 642                                       rscp->cred.vc_uid);
 643                         rsc_put(&rscp->h, &rsc_cache);
 644                         rsc_cache.entries--;
 645                 }
 646         }
 647         write_unlock(&rsc_cache.hash_lock);
 648         EXIT;
 649 }
 650
 651 static struct cache_detail rsc_cache = {
 652         .hash_size      = RSC_HASHMAX,
 653         .hash_table     = rsc_table,
 654         .name           = "auth.ptlrpcs.context",
 655         .cache_put      = rsc_put,
 656         .cache_parse    = rsc_parse,
 657 };
 658
 659 static struct rsc *
 660 gss_svc_searchbyctx(rawobj_t *handle)
 661 {
 662         struct rsc rsci;
 663         struct rsc *found;
 664
 665         rsci.handle = *handle;
 666         found = rsc_lookup(&rsci, 0);
 667         if (!found)
 668                 return NULL;
 669
 670         if (cache_check(&rsc_cache, &found->h, NULL))
 671                 return NULL;
 672
 673         return found;
 674 }
 675
 676 /* FIXME
 677  * again hacking: only try to give the svcgssd a chance to handle
 678  * upcalls.
 679  */
 680 struct cache_deferred_req* my_defer(struct cache_req *req)
 681 {
 682         yield();
 683         return NULL;
 684 }
 685 static struct cache_req my_chandle = {my_defer};
 686
 687 /* Implements sequence number algorithm as specified in RFC 2203. */
 688 static int
 689 gss_check_seq_num(struct gss_svc_seq_data *sd, __u32 seq_num)
 690 {
 691         int rc = 0;
 692
 693         spin_lock(&sd->sd_lock);
 694         if (seq_num > sd->sd_max) {
 695                 if (seq_num >= sd->sd_max + GSS_SEQ_WIN) {
 696                         memset(sd->sd_win, 0, sizeof(sd->sd_win));
 697                         sd->sd_max = seq_num;
 698                 } else {
 699                         while(sd->sd_max < seq_num) {
 700                                 sd->sd_max++;
 701                                 __clear_bit(sd->sd_max % GSS_SEQ_WIN,
 702                                             sd->sd_win);
 703                         }
 704                 }
 705                 __set_bit(seq_num % GSS_SEQ_WIN, sd->sd_win);
 706                 goto exit;
 707         } else if (seq_num + GSS_SEQ_WIN <= sd->sd_max) {
 708                 CERROR("seq %u too low: max %u, win %d\n",
 709                         seq_num, sd->sd_max, GSS_SEQ_WIN);
 710                 rc = 1;
 711                 goto exit;
 712         }
 713
 714         if (__test_and_set_bit(seq_num % GSS_SEQ_WIN, sd->sd_win)) {
 715                 CERROR("seq %u is replay: max %u, win %d\n",
 716                         seq_num, sd->sd_max, GSS_SEQ_WIN);
 717                 rc = 1;
 718         }
 719 exit:
 720         spin_unlock(&sd->sd_lock);
 721         return rc;
 722 }
 723
 724 static int
 725 gss_svc_verify_request(struct ptlrpc_request *req,
 726                        struct rsc *rsci,
 727                        struct rpc_gss_wire_cred *gc,
 728                        __u32 *vp, __u32 vlen)
 729 {
 730         struct ptlrpcs_wire_hdr *sec_hdr;
 731         struct gss_ctx *ctx = rsci->mechctx;
 732         __u32 maj_stat;
 733         rawobj_t msg;
 734         rawobj_t mic;
 735         ENTRY;
 736
 737         sec_hdr = (struct ptlrpcs_wire_hdr *) req->rq_reqbuf;
 738
 739         req->rq_reqmsg = (struct lustre_msg *) (req->rq_reqbuf + sizeof(*sec_hdr));
 740         req->rq_reqlen = sec_hdr->msg_len;
 741
 742         msg.len = sec_hdr->msg_len;
 743         msg.data = (__u8 *)req->rq_reqmsg;
 744
 745         mic.len = le32_to_cpu(*vp++);
 746         mic.data = (unsigned char *)vp;
 747         vlen -= 4;
 748
 749         if (mic.len > vlen) {
 750                 CERROR("checksum len %d, while buffer len %d\n",
 751                         mic.len, vlen);
 752                 RETURN(GSS_S_CALL_BAD_STRUCTURE);
 753         }
 754
 755         if (mic.len > 256) {
 756                 CERROR("invalid mic len %d\n", mic.len);
 757                 RETURN(GSS_S_CALL_BAD_STRUCTURE);
 758         }
 759
 760         maj_stat = kgss_verify_mic(ctx, &msg, &mic, NULL);
 761         if (maj_stat != GSS_S_COMPLETE) {
 762                 CERROR("MIC verification error: major %x\n", maj_stat);
 763                 RETURN(maj_stat);
 764         }
 765
 766         if (gss_check_seq_num(&rsci->seqdata, gc->gc_seq)) {
 767                 CERROR("discard replayed request %p(o%u,x"LPU64",t"LPU64")\n",
 768                         req, req->rq_reqmsg->opc, req->rq_xid,
 769                         req->rq_reqmsg->transno);
 770                 RETURN(GSS_S_DUPLICATE_TOKEN);
 771         }
 772
 773         RETURN(GSS_S_COMPLETE);
 774 }
 775
 776 static int
 777 gss_svc_unseal_request(struct ptlrpc_request *req,
 778                        struct rsc *rsci,
 779                        struct rpc_gss_wire_cred *gc,
 780                        __u32 *vp, __u32 vlen)
 781 {
 782         struct ptlrpcs_wire_hdr *sec_hdr;
 783         struct gss_ctx *ctx = rsci->mechctx;
 784         rawobj_t cipher_text, plain_text;
 785         __u32 major;
 786         ENTRY;
 787
 788         sec_hdr = (struct ptlrpcs_wire_hdr *) req->rq_reqbuf;
 789
 790         if (vlen < 4) {
 791                 CERROR("vlen only %u\n", vlen);
 792                 RETURN(GSS_S_CALL_BAD_STRUCTURE);
 793         }
 794
 795         cipher_text.len = le32_to_cpu(*vp++);
 796         cipher_text.data = (__u8 *) vp;
 797         vlen -= 4;
 798
 799         if (cipher_text.len > vlen) {
 800                 CERROR("cipher claimed %u while buf only %u\n",
 801                         cipher_text.len, vlen);
 802                 RETURN(GSS_S_CALL_BAD_STRUCTURE);
 803         }
 804
 805         plain_text = cipher_text;
 806
 807         major = kgss_unwrap(ctx, GSS_C_QOP_DEFAULT, &cipher_text, &plain_text);
 808         if (major) {
 809                 CERROR("unwrap error 0x%x\n", major);
 810                 RETURN(major);
 811         }
 812
 813         if (gss_check_seq_num(&rsci->seqdata, gc->gc_seq)) {
 814                 CERROR("discard replayed request %p(o%u,x"LPU64",t"LPU64")\n",
 815                         req, req->rq_reqmsg->opc, req->rq_xid,
 816                         req->rq_reqmsg->transno);
 817                 RETURN(GSS_S_DUPLICATE_TOKEN);
 818         }
 819
 820         req->rq_reqmsg = (struct lustre_msg *) (vp);
 821         req->rq_reqlen = plain_text.len;
 822
 823         CDEBUG(D_SEC, "msg len %d\n", req->rq_reqlen);
 824
 825         RETURN(GSS_S_COMPLETE);
 826 }
 827
 828 static int
 829 gss_pack_err_notify(struct ptlrpc_request *req,
 830                     __u32 major, __u32 minor)
 831 {
 832         struct gss_svc_data *svcdata = req->rq_svcsec_data;
 833         __u32 reslen, *resp, *reslenp;
 834         char  nidstr[PTL_NALFMT_SIZE];
 835         const __u32 secdata_len = 7 * 4;
 836         int rc;
 837         ENTRY;
 838
 839         OBD_FAIL_RETURN(OBD_FAIL_SVCGSS_ERR_NOTIFY|OBD_FAIL_ONCE, -EINVAL);
 840
 841         LASSERT(svcdata);
 842         svcdata->is_err_notify = 1;
 843         svcdata->reserve_len = 7 * 4;
 844
 845         rc = lustre_pack_reply(req, 0, NULL, NULL);
 846         if (rc) {
 847                 CERROR("could not pack reply, err %d\n", rc);
 848                 RETURN(rc);
 849         }
 850
 851         LASSERT(req->rq_reply_state);
 852         LASSERT(req->rq_reply_state->rs_repbuf);
 853         LASSERT(req->rq_reply_state->rs_repbuf_len >= secdata_len);
 854         resp = (__u32 *) req->rq_reply_state->rs_repbuf;
 855
 856         /* header */
 857         *resp++ = cpu_to_le32(PTLRPCS_FLVR_GSS_NONE);
 858         *resp++ = cpu_to_le32(PTLRPCS_SVC_NONE);
 859         *resp++ = cpu_to_le32(req->rq_replen);
 860         reslenp = resp++;
 861
 862         /* skip lustre msg */
 863         resp += req->rq_replen / 4;
 864         reslen = svcdata->reserve_len;
 865
 866         /* gss replay:
 867          * version, subflavor, notify, major, minor,
 868          * obj1(fake), obj2(fake)
 869          */
 870         *resp++ = cpu_to_le32(PTLRPC_SEC_GSS_VERSION);
 871         *resp++ = cpu_to_le32(PTLRPCS_FLVR_KRB5I);
 872         *resp++ = cpu_to_le32(PTLRPCS_GSS_PROC_ERR);
 873         *resp++ = cpu_to_le32(major);
 874         *resp++ = cpu_to_le32(minor);
 875         *resp++ = 0;
 876         *resp++ = 0;
 877         reslen -= (4 * 4);
 878         /* the actual sec data length */
 879         *reslenp = cpu_to_le32(secdata_len);
 880
 881         req->rq_reply_state->rs_repdata_len += (secdata_len);
 882         CDEBUG(D_SEC, "prepare gss error notify(0x%x/0x%x) to %s\n",
 883                major, minor,
 884                portals_nid2str(req->rq_peer.peer_ni->pni_number,
 885                                req->rq_peer.peer_id.nid, nidstr));
 886         RETURN(0);
 887 }
 888
 889 static void dump_cache_head(struct cache_head *h)
 890 {
 891         CWARN("ref %d, fl %lx, n %p, t %ld, %ld\n",
 892               atomic_read(&h->refcnt), h->flags, h->next,
 893               h->expiry_time, h->last_refresh);
 894 }
 895 static void dump_rsi(struct rsi *rsi)
 896 {
 897         CWARN("dump rsi %p\n", rsi);
 898         dump_cache_head(&rsi->h);
 899         CWARN("%x,%x,%llx\n", rsi->naltype, rsi->netid, rsi->nid);
 900         CWARN("len %d, d %p\n", rsi->in_handle.len, rsi->in_handle.data);
 901         CWARN("len %d, d %p\n", rsi->in_token.len, rsi->in_token.data);
 902         CWARN("len %d, d %p\n", rsi->out_handle.len, rsi->out_handle.data);
 903         CWARN("len %d, d %p\n", rsi->out_token.len, rsi->out_token.data);
 904 }
 905
 906 static int
 907 gss_svcsec_handle_init(struct ptlrpc_request *req,
 908                        struct rpc_gss_wire_cred *gc,
 909                        __u32 *secdata, __u32 seclen,
 910                        enum ptlrpcs_error *res)
 911 {
 912         struct gss_svc_data *svcdata = req->rq_svcsec_data;
 913         struct rsc          *rsci;
 914         struct rsi          *rsikey, *rsip;
 915         rawobj_t             tmpobj;
 916         __u32 reslen,       *resp, *reslenp;
 917         char                 nidstr[PTL_NALFMT_SIZE];
 918         int                  rc;
 919         ENTRY;
 920
 921         LASSERT(svcdata);
 922
 923         CDEBUG(D_SEC, "processing gss init(%d) request from %s\n", gc->gc_proc,
 924                portals_nid2str(req->rq_peer.peer_ni->pni_number,
 925                                req->rq_peer.peer_id.nid, nidstr));
 926
 927         *res = PTLRPCS_BADCRED;
 928         OBD_FAIL_RETURN(OBD_FAIL_SVCGSS_INIT_REQ|OBD_FAIL_ONCE, SVC_DROP);
 929
 930         if (gc->gc_proc == RPC_GSS_PROC_INIT &&
 931             gc->gc_ctx.len != 0) {
 932                 CERROR("proc %d, ctx_len %d: not really init?\n",
 933                 gc->gc_proc == RPC_GSS_PROC_INIT, gc->gc_ctx.len);
 934                 RETURN(SVC_DROP);
 935         }
 936
 937         OBD_ALLOC(rsikey, sizeof(*rsikey));
 938         if (!rsikey) {
 939                 CERROR("out of memory\n");
 940                 RETURN(SVC_DROP);
 941         }
 942         cache_init(&rsikey->h);
 943
 944         /* obtain lustre svc type */
 945         if (seclen < 4) {
 946                 CERROR("sec size %d too small\n", seclen);
 947                 GOTO(out_rsikey, rc = SVC_DROP);
 948         }
 949         rsikey->lustre_svc = le32_to_cpu(*secdata++);
 950         seclen -= 4;
 951
 952         /* duplicate context handle. currently always 0 */
 953         if (rawobj_dup(&rsikey->in_handle, &gc->gc_ctx)) {
 954                 CERROR("fail to dup context handle\n");
 955                 GOTO(out_rsikey, rc = SVC_DROP);
 956         }
 957
 958         /* extract token */
 959         *res = PTLRPCS_BADVERF;
 960         if (rawobj_extract(&tmpobj, &secdata, &seclen)) {
 961                 CERROR("can't extract token\n");
 962                 GOTO(out_rsikey, rc = SVC_DROP);
 963         }
 964         if (rawobj_dup(&rsikey->in_token, &tmpobj)) {
 965                 CERROR("can't duplicate token\n");
 966                 GOTO(out_rsikey, rc = SVC_DROP);
 967         }
 968
 969         rsikey->naltype = (__u32) req->rq_peer.peer_ni->pni_number;
 970         rsikey->netid = 0;
 971         rsikey->nid = (__u64) req->rq_peer.peer_id.nid;
 972
 973         rsip = gssd_upcall(rsikey, &my_chandle);
 974         if (!rsip) {
 975                 CERROR("error in gssd_upcall.\n");
 976
 977                 rc = SVC_COMPLETE;
 978                 if (gss_pack_err_notify(req, GSS_S_FAILURE, 0))
 979                         rc = SVC_DROP;
 980
 981                 GOTO(out_rsikey, rc);
 982         }
 983
 984         rsci = gss_svc_searchbyctx(&rsip->out_handle);
 985         if (!rsci) {
 986                 CERROR("rsci still not mature yet?\n");
 987
 988                 rc = SVC_COMPLETE;
 989                 if (gss_pack_err_notify(req, GSS_S_FAILURE, 0))
 990                         rc = SVC_DROP;
 991
 992                 GOTO(out_rsip, rc);
 993         }
 994         CDEBUG(D_SEC, "svcsec create gss context %p(%u@%s)\n",
 995                rsci, rsci->cred.vc_uid,
 996                portals_nid2str(req->rq_peer.peer_ni->pni_number,
 997                                req->rq_peer.peer_id.nid, nidstr));
 998
 999         svcdata->is_init = 1;
1000         svcdata->reserve_len = 6 * 4 +
1001                 size_round4(rsip->out_handle.len) +
1002                 size_round4(rsip->out_token.len);
1003
1004         rc = lustre_pack_reply(req, 0, NULL, NULL);
1005         if (rc) {
1006                 CERROR("failed to pack reply, rc = %d\n", rc);
1007                 set_bit(CACHE_NEGATIVE, &rsci->h.flags);
1008                 GOTO(out, rc = SVC_DROP);
1009         }
1010
1011         /* header */
1012         resp = (__u32 *) req->rq_reply_state->rs_repbuf;
1013         *resp++ = cpu_to_le32(PTLRPCS_FLVR_GSS_NONE);
1014         *resp++ = cpu_to_le32(PTLRPCS_SVC_NONE);
1015         *resp++ = cpu_to_le32(req->rq_replen);
1016         reslenp = resp++;
1017
1018         resp += req->rq_replen / 4;
1019         reslen = svcdata->reserve_len;
1020
1021         /* gss reply:
1022          * status, major, minor, seq, out_handle, out_token
1023          */
1024         *resp++ = cpu_to_le32(PTLRPCS_OK);
1025         *resp++ = cpu_to_le32(rsip->major_status);
1026         *resp++ = cpu_to_le32(rsip->minor_status);
1027         *resp++ = cpu_to_le32(GSS_SEQ_WIN);
1028         reslen -= (4 * 4);
1029         if (rawobj_serialize(&rsip->out_handle,
1030                              &resp, &reslen)) {
1031                 dump_rsi(rsip);
1032                 dump_rsi(rsikey);
1033                 LBUG();
1034         }
1035         if (rawobj_serialize(&rsip->out_token,
1036                              &resp, &reslen)) {
1037                 dump_rsi(rsip);
1038                 dump_rsi(rsikey);
1039                 LBUG();
1040         }
1041         /* the actual sec data length */
1042         *reslenp = cpu_to_le32(svcdata->reserve_len - reslen);
1043
1044         req->rq_reply_state->rs_repdata_len += le32_to_cpu(*reslenp);
1045         CDEBUG(D_SEC, "req %p: msgsize %d, authsize %d, "
1046                "total size %d\n", req, req->rq_replen,
1047                le32_to_cpu(*reslenp),
1048                req->rq_reply_state->rs_repdata_len);
1049
1050         *res = PTLRPCS_OK;
1051
1052         req->rq_remote_realm = rsci->remote_realm;
1053         req->rq_auth_usr_mds = rsci->auth_usr_mds;
1054         req->rq_auth_usr_oss = rsci->auth_usr_oss;
1055         req->rq_auth_uid = rsci->cred.vc_uid;
1056         req->rq_mapped_uid = rsci->mapped_uid;
1057
1058         if (req->rq_auth_usr_mds) {
1059                 CWARN("usr from %s authenticated as mds svc cred\n",
1060                 portals_nid2str(req->rq_peer.peer_ni->pni_number,
1061                                 req->rq_peer.peer_id.nid, nidstr));
1062         }
1063         if (req->rq_auth_usr_oss) {
1064                 CWARN("usr from %s authenticated as oss svc cred\n",
1065                 portals_nid2str(req->rq_peer.peer_ni->pni_number,
1066                                 req->rq_peer.peer_id.nid, nidstr));
1067         }
1068
1069         /* This is simplified since right now we doesn't support
1070          * INIT_CONTINUE yet.
1071          */
1072         if (gc->gc_proc == RPC_GSS_PROC_INIT) {
1073                 struct ptlrpcs_wire_hdr *hdr;
1074
1075                 hdr = buf_to_sec_hdr(req->rq_reqbuf);
1076                 req->rq_reqmsg = buf_to_lustre_msg(req->rq_reqbuf);
1077                 req->rq_reqlen = hdr->msg_len;
1078
1079                 rc = SVC_LOGIN;
1080         } else
1081                 rc = SVC_COMPLETE;
1082
1083 out:
1084         rsc_put(&rsci->h, &rsc_cache);
1085 out_rsip:
1086         rsi_put(&rsip->h, &rsi_cache);
1087 out_rsikey:
1088         rsi_put(&rsikey->h, &rsi_cache);
1089
1090         RETURN(rc);
1091 }
1092
1093 static int
1094 gss_svcsec_handle_data(struct ptlrpc_request *req,
1095                        struct rpc_gss_wire_cred *gc,
1096                        __u32 *secdata, __u32 seclen,
1097                        enum ptlrpcs_error *res)
1098 {
1099         struct rsc          *rsci;
1100         char                 nidstr[PTL_NALFMT_SIZE];
1101         __u32                major;
1102         int                  rc;
1103         ENTRY;
1104
1105         *res = PTLRPCS_GSS_CREDPROBLEM;
1106
1107         rsci = gss_svc_searchbyctx(&gc->gc_ctx);
1108         if (!rsci) {
1109                 CWARN("Invalid gss context handle from %s\n",
1110                        portals_nid2str(req->rq_peer.peer_ni->pni_number,
1111                                        req->rq_peer.peer_id.nid, nidstr));
1112                 major = GSS_S_NO_CONTEXT;
1113                 goto notify_err;
1114         }
1115
1116         switch (gc->gc_svc) {
1117         case PTLRPCS_GSS_SVC_INTEGRITY:
1118                 major = gss_svc_verify_request(req, rsci, gc, secdata, seclen);
1119                 if (major == GSS_S_COMPLETE)
1120                         break;
1121
1122                 CWARN("fail in verify:0x%x: ctx %p@%s\n", major, rsci,
1123                        portals_nid2str(req->rq_peer.peer_ni->pni_number,
1124                                        req->rq_peer.peer_id.nid, nidstr));
1125                 goto notify_err;
1126         case PTLRPCS_GSS_SVC_PRIVACY:
1127                 major = gss_svc_unseal_request(req, rsci, gc, secdata, seclen);
1128                 if (major == GSS_S_COMPLETE)
1129                         break;
1130
1131                 CWARN("fail in decrypt:0x%x: ctx %p@%s\n", major, rsci,
1132                        portals_nid2str(req->rq_peer.peer_ni->pni_number,
1133                                        req->rq_peer.peer_id.nid, nidstr));
1134                 goto notify_err;
1135         default:
1136                 CERROR("unsupported gss service %d\n", gc->gc_svc);
1137                 GOTO(out, rc = SVC_DROP);
1138         }
1139
1140         req->rq_remote_realm = rsci->remote_realm;
1141         req->rq_auth_usr_mds = rsci->auth_usr_mds;
1142         req->rq_auth_usr_oss = rsci->auth_usr_oss;
1143         req->rq_auth_uid = rsci->cred.vc_uid;
1144         req->rq_mapped_uid = rsci->mapped_uid;
1145
1146         *res = PTLRPCS_OK;
1147         GOTO(out, rc = SVC_OK);
1148
1149 notify_err:
1150         if (gss_pack_err_notify(req, major, 0))
1151                 rc = SVC_DROP;
1152         else
1153                 rc = SVC_COMPLETE;
1154 out:
1155         if (rsci)
1156                 rsc_put(&rsci->h, &rsc_cache);
1157         RETURN(rc);
1158 }
1159
1160 static int
1161 gss_svcsec_handle_destroy(struct ptlrpc_request *req,
1162                           struct rpc_gss_wire_cred *gc,
1163                           __u32 *secdata, __u32 seclen,
1164                           enum ptlrpcs_error *res)
1165 {
1166         struct gss_svc_data *svcdata = req->rq_svcsec_data;
1167         struct rsc          *rsci;
1168         char                 nidstr[PTL_NALFMT_SIZE];
1169         int                  rc;
1170         ENTRY;
1171
1172         LASSERT(svcdata);
1173         *res = PTLRPCS_GSS_CREDPROBLEM;
1174
1175         rsci = gss_svc_searchbyctx(&gc->gc_ctx);
1176         if (!rsci) {
1177                 CWARN("invalid gss context handle for destroy.\n");
1178                 RETURN(SVC_DROP);
1179         }
1180
1181         if (gc->gc_svc != PTLRPCS_GSS_SVC_INTEGRITY) {
1182                 CERROR("service %d is not supported in destroy.\n",
1183                         gc->gc_svc);
1184                 GOTO(out, rc = SVC_DROP);
1185         }
1186
1187         *res = gss_svc_verify_request(req, rsci, gc, secdata, seclen);
1188         if (*res)
1189                 GOTO(out, rc = SVC_DROP);
1190
1191         /* compose reply, which is actually nothing */
1192         svcdata->is_fini = 1;
1193         if (lustre_pack_reply(req, 0, NULL, NULL))
1194                 GOTO(out, rc = SVC_DROP);
1195
1196         CDEBUG(D_SEC, "svcsec destroy gss context %p(%u@%s)\n",
1197                rsci, rsci->cred.vc_uid,
1198                portals_nid2str(req->rq_peer.peer_ni->pni_number,
1199                                req->rq_peer.peer_id.nid, nidstr));
1200
1201         set_bit(CACHE_NEGATIVE, &rsci->h.flags);
1202         *res = PTLRPCS_OK;
1203         rc = SVC_LOGOUT;
1204 out:
1205         rsc_put(&rsci->h, &rsc_cache);
1206         RETURN(rc);
1207 }
1208
1209 /*
1210  * let incomming request go through security check:
1211  *  o context establishment: invoke user space helper
1212  *  o data exchange: verify/decrypt
1213  *  o context destruction: mark context invalid
1214  *
1215  * in most cases, error will result to drop the packet silently.
1216  */
1217 static int
1218 gss_svcsec_accept(struct ptlrpc_request *req, enum ptlrpcs_error *res)
1219 {
1220         struct gss_svc_data *svcdata;
1221         struct rpc_gss_wire_cred *gc;
1222         struct ptlrpcs_wire_hdr *sec_hdr;
1223         __u32 subflavor, seclen, *secdata, version;
1224         int rc;
1225         ENTRY;
1226
1227         CDEBUG(D_SEC, "request %p\n", req);
1228         LASSERT(req->rq_reqbuf);
1229         LASSERT(req->rq_reqbuf_len);
1230
1231         *res = PTLRPCS_BADCRED;
1232
1233         sec_hdr = buf_to_sec_hdr(req->rq_reqbuf);
1234         LASSERT(SEC_FLAVOR_MAJOR(sec_hdr->flavor) == PTLRPCS_FLVR_MAJOR_GSS);
1235
1236         seclen = req->rq_reqbuf_len - sizeof(*sec_hdr) - sec_hdr->msg_len;
1237         secdata = (__u32 *) buf_to_sec_data(req->rq_reqbuf);
1238
1239         if (sec_hdr->sec_len > seclen) {
1240                 CERROR("seclen %d, while max buf %d\n",
1241                         sec_hdr->sec_len, seclen);
1242                 RETURN(SVC_DROP);
1243         }
1244
1245         if (seclen < 6 * 4) {
1246                 CERROR("sec size %d too small\n", seclen);
1247                 RETURN(SVC_DROP);
1248         }
1249
1250         LASSERT(!req->rq_svcsec_data);
1251         OBD_ALLOC(svcdata, sizeof(*svcdata));
1252         if (!svcdata) {
1253                 CERROR("fail to alloc svcdata\n");
1254                 RETURN(SVC_DROP);
1255         }
1256         req->rq_svcsec_data = svcdata;
1257         gc = &svcdata->clcred;
1258
1259         /* Now secdata/seclen is what we want to parse
1260          */
1261         version = le32_to_cpu(*secdata++);      /* version */
1262         subflavor = le32_to_cpu(*secdata++);    /* subflavor */
1263         gc->gc_proc = le32_to_cpu(*secdata++);  /* proc */
1264         gc->gc_seq = le32_to_cpu(*secdata++);   /* seq */
1265         gc->gc_svc = le32_to_cpu(*secdata++);   /* service */
1266         seclen -= 5 * 4;
1267
1268         CDEBUG(D_SEC, "wire gss_hdr: %u/%u/%u/%u/%u\n",
1269                version, subflavor, gc->gc_proc,
1270                gc->gc_seq, gc->gc_svc);
1271
1272         if (version != PTLRPC_SEC_GSS_VERSION) {
1273                 CERROR("gss version mismatch: %d - %d\n",
1274                         version, PTLRPC_SEC_GSS_VERSION);
1275                 GOTO(err_free, rc = SVC_DROP);
1276         }
1277
1278         /* We _must_ alloc new storage for gc_ctx. In case of recovery
1279          * request will be saved to delayed handling, at that time the
1280          * incoming buffer might have already been released.
1281          */
1282         if (rawobj_extract_alloc(&gc->gc_ctx, &secdata, &seclen)) {
1283                 CERROR("fail to obtain gss context handle\n");
1284                 GOTO(err_free, rc = SVC_DROP);
1285         }
1286
1287         *res = PTLRPCS_BADVERF;
1288         switch(gc->gc_proc) {
1289         case RPC_GSS_PROC_INIT:
1290         case RPC_GSS_PROC_CONTINUE_INIT:
1291                 rc = gss_svcsec_handle_init(req, gc, secdata, seclen, res);
1292                 break;
1293         case RPC_GSS_PROC_DATA:
1294                 rc = gss_svcsec_handle_data(req, gc, secdata, seclen, res);
1295                 break;
1296         case RPC_GSS_PROC_DESTROY:
1297                 rc = gss_svcsec_handle_destroy(req, gc, secdata, seclen, res);
1298                 break;
1299         default:
1300                 rc = SVC_DROP;
1301                 LBUG();
1302         }
1303
1304 err_free:
1305         if (rc == SVC_DROP && req->rq_svcsec_data) {
1306                 OBD_FREE(req->rq_svcsec_data, sizeof(struct gss_svc_data));
1307                 req->rq_svcsec_data = NULL;
1308         }
1309
1310         RETURN(rc);
1311 }
1312
1313 static int
1314 gss_svcsec_authorize(struct ptlrpc_request *req)
1315 {
1316         struct ptlrpc_reply_state *rs = req->rq_reply_state;
1317         struct gss_svc_data *gsd = (struct gss_svc_data *)req->rq_svcsec_data;
1318         struct rpc_gss_wire_cred  *gc = &gsd->clcred;
1319         struct rsc                *rscp;
1320         struct ptlrpcs_wire_hdr   *sec_hdr;
1321         rawobj_buf_t               msg_buf;
1322         rawobj_t                   cipher_buf;
1323         __u32                     *vp, *vpsave, major, vlen, seclen;
1324         rawobj_t                   lmsg, mic;
1325         int                        ret;
1326         ENTRY;
1327
1328         LASSERT(rs);
1329         LASSERT(rs->rs_repbuf);
1330         LASSERT(gsd);
1331
1332         if (gsd->is_init || gsd->is_init_continue ||
1333             gsd->is_err_notify || gsd->is_fini) {
1334                 /* nothing to do in these cases */
1335                 CDEBUG(D_SEC, "req %p: init/fini/err\n", req);
1336                 RETURN(0);
1337         }
1338
1339         if (gc->gc_proc != RPC_GSS_PROC_DATA) {
1340                 CERROR("proc %d not support\n", gc->gc_proc);
1341                 RETURN(-EINVAL);
1342         }
1343
1344         rscp = gss_svc_searchbyctx(&gc->gc_ctx);
1345         if (!rscp) {
1346                 CERROR("ctx %u disapeared under us\n",
1347                        *((__u32 *) gc->gc_ctx.data));
1348                 RETURN(-EINVAL);
1349         }
1350
1351         sec_hdr = (struct ptlrpcs_wire_hdr *) rs->rs_repbuf;
1352         switch (gc->gc_svc) {
1353         case  PTLRPCS_GSS_SVC_INTEGRITY:
1354                 /* prepare various pointers */
1355                 lmsg.len = req->rq_replen;
1356                 lmsg.data = (__u8 *) (rs->rs_repbuf + sizeof(*sec_hdr));
1357                 vp = (__u32 *) (lmsg.data + lmsg.len);
1358                 vlen = rs->rs_repbuf_len - sizeof(*sec_hdr) - lmsg.len;
1359                 seclen = vlen;
1360
1361                 sec_hdr->flavor = cpu_to_le32(PTLRPCS_FLVR_GSS_AUTH);
1362                 sec_hdr->msg_len = cpu_to_le32(req->rq_replen);
1363
1364                 /* standard gss hdr */
1365                 LASSERT(vlen >= 7 * 4);
1366                 *vp++ = cpu_to_le32(PTLRPC_SEC_GSS_VERSION);
1367                 *vp++ = cpu_to_le32(PTLRPCS_FLVR_KRB5I);
1368                 *vp++ = cpu_to_le32(RPC_GSS_PROC_DATA);
1369                 *vp++ = cpu_to_le32(gc->gc_seq);
1370                 *vp++ = cpu_to_le32(PTLRPCS_GSS_SVC_INTEGRITY);
1371                 *vp++ = 0;      /* fake ctx handle */
1372                 vpsave = vp++;  /* reserve size */
1373                 vlen -= 7 * 4;
1374
1375                 mic.len = vlen;
1376                 mic.data = (unsigned char *)vp;
1377
1378                 major = kgss_get_mic(rscp->mechctx, 0, &lmsg, &mic);
1379                 if (major) {
1380                         CERROR("fail to get MIC: 0x%x\n", major);
1381                         GOTO(out, ret = -EINVAL);
1382                 }
1383                 *vpsave = cpu_to_le32(mic.len);
1384                 seclen = seclen - vlen + mic.len;
1385                 sec_hdr->sec_len = cpu_to_le32(seclen);
1386                 rs->rs_repdata_len += size_round(seclen);
1387                 break;
1388         case  PTLRPCS_GSS_SVC_PRIVACY:
1389                 vp = (__u32 *) (rs->rs_repbuf + sizeof(*sec_hdr));
1390                 vlen = rs->rs_repbuf_len - sizeof(*sec_hdr);
1391                 seclen = vlen;
1392
1393                 sec_hdr->flavor = cpu_to_le32(PTLRPCS_FLVR_GSS_PRIV);
1394                 sec_hdr->msg_len = cpu_to_le32(0);
1395
1396                 /* standard gss hdr */
1397                 LASSERT(vlen >= 7 * 4);
1398                 *vp++ = cpu_to_le32(PTLRPC_SEC_GSS_VERSION);
1399                 *vp++ = cpu_to_le32(PTLRPCS_FLVR_KRB5I);
1400                 *vp++ = cpu_to_le32(RPC_GSS_PROC_DATA);
1401                 *vp++ = cpu_to_le32(gc->gc_seq);
1402                 *vp++ = cpu_to_le32(PTLRPCS_GSS_SVC_PRIVACY);
1403                 *vp++ = 0;      /* fake ctx handle */
1404                 vpsave = vp++;  /* reserve size */
1405                 vlen -= 7 * 4;
1406
1407                 msg_buf.buf = (__u8 *) rs->rs_msg - GSS_PRIVBUF_PREFIX_LEN;
1408                 msg_buf.buflen = req->rq_replen + GSS_PRIVBUF_PREFIX_LEN +
1409                                  GSS_PRIVBUF_SUFFIX_LEN;
1410                 msg_buf.dataoff = GSS_PRIVBUF_PREFIX_LEN;
1411                 msg_buf.datalen = req->rq_replen;
1412
1413                 cipher_buf.data = (__u8 *) vp;
1414                 cipher_buf.len = vlen;
1415
1416                 major = kgss_wrap(rscp->mechctx, GSS_C_QOP_DEFAULT,
1417                                 &msg_buf, &cipher_buf);
1418                 if (major) {
1419                         CERROR("failed to wrap: 0x%x\n", major);
1420                         GOTO(out, ret = -EINVAL);
1421                 }
1422
1423                 *vpsave = cpu_to_le32(cipher_buf.len);
1424                 seclen = seclen - vlen + cipher_buf.len;
1425                 sec_hdr->sec_len = cpu_to_le32(seclen);
1426                 rs->rs_repdata_len += size_round(seclen);
1427                 break;
1428         default:
1429                 CERROR("Unknown service %d\n", gc->gc_svc);
1430                 GOTO(out, ret = -EINVAL);
1431         }
1432         ret = 0;
1433 out:
1434         rsc_put(&rscp->h, &rsc_cache);
1435
1436         RETURN(ret);
1437 }
1438
1439 static
1440 void gss_svcsec_cleanup_req(struct ptlrpc_svcsec *svcsec,
1441                             struct ptlrpc_request *req)
1442 {
1443         struct gss_svc_data *gsd = (struct gss_svc_data *) req->rq_svcsec_data;
1444
1445         if (!gsd) {
1446                 CDEBUG(D_SEC, "no svc_data present. do nothing\n");
1447                 return;
1448         }
1449
1450         /* gc_ctx is allocated, see gss_svcsec_accept() */
1451         rawobj_free(&gsd->clcred.gc_ctx);
1452
1453         OBD_FREE(gsd, sizeof(*gsd));
1454         req->rq_svcsec_data = NULL;
1455         return;
1456 }
1457
1458 static
1459 int gss_svcsec_est_payload(struct ptlrpc_svcsec *svcsec,
1460                            struct ptlrpc_request *req,
1461                            int msgsize)
1462 {
1463         struct gss_svc_data *svcdata = req->rq_svcsec_data;
1464         ENTRY;
1465
1466         /* just return the pre-set reserve_len for init/fini/err cases.
1467          */
1468         LASSERT(svcdata);
1469         if (svcdata->is_init) {
1470                 CDEBUG(D_SEC, "is_init, reserver size %d(%d)\n",
1471                        size_round(svcdata->reserve_len),
1472                        svcdata->reserve_len);
1473                 LASSERT(svcdata->reserve_len);
1474                 LASSERT(svcdata->reserve_len % 4 == 0);
1475                 RETURN(size_round(svcdata->reserve_len));
1476         } else if (svcdata->is_err_notify) {
1477                 CDEBUG(D_SEC, "is_err_notify, reserver size %d(%d)\n",
1478                        size_round(svcdata->reserve_len),
1479                        svcdata->reserve_len);
1480                 RETURN(size_round(svcdata->reserve_len));
1481         } else if (svcdata->is_fini) {
1482                 CDEBUG(D_SEC, "is_fini, reserver size 0\n");
1483                 RETURN(0);
1484         } else {
1485                 if (svcdata->clcred.gc_svc == PTLRPCS_GSS_SVC_NONE ||
1486                     svcdata->clcred.gc_svc == PTLRPCS_GSS_SVC_INTEGRITY)
1487                         RETURN(size_round(GSS_MAX_AUTH_PAYLOAD));
1488                 else if (svcdata->clcred.gc_svc == PTLRPCS_GSS_SVC_PRIVACY)
1489                         RETURN(size_round16(GSS_MAX_AUTH_PAYLOAD + msgsize +
1490                                             GSS_PRIVBUF_PREFIX_LEN +
1491                                             GSS_PRIVBUF_SUFFIX_LEN));
1492                 else {
1493                         CERROR("unknown gss svc %u\n", svcdata->clcred.gc_svc);
1494                         *((int *)0) = 0;
1495                         LBUG();
1496                 }
1497         }
1498         RETURN(0);
1499 }
1500
1501 int gss_svcsec_alloc_repbuf(struct ptlrpc_svcsec *svcsec,
1502                             struct ptlrpc_request *req,
1503                             int msgsize)
1504 {
1505         struct gss_svc_data *gsd = (struct gss_svc_data *) req->rq_svcsec_data;
1506         struct ptlrpc_reply_state *rs;
1507         int msg_payload, sec_payload;
1508         int privacy, rc;
1509         ENTRY;
1510
1511         /* determine the security type: none/auth or priv, we have
1512          * different pack scheme for them.
1513          * init/fini/err will always be treated as none/auth.
1514          */
1515         LASSERT(gsd);
1516         if (!gsd->is_init && !gsd->is_init_continue &&
1517             !gsd->is_fini && !gsd->is_err_notify &&
1518             gsd->clcred.gc_svc == PTLRPCS_GSS_SVC_PRIVACY)
1519                 privacy = 1;
1520         else
1521                 privacy = 0;
1522
1523         msg_payload = privacy ? 0 : msgsize;
1524         sec_payload = gss_svcsec_est_payload(svcsec, req, msgsize);
1525
1526         rc = svcsec_alloc_reply_state(req, msg_payload, sec_payload);
1527         if (rc)
1528                 RETURN(rc);
1529
1530         rs = req->rq_reply_state;
1531         LASSERT(rs);
1532         rs->rs_msg_len = msgsize;
1533
1534         if (privacy) {
1535                 /* we can choose to let msg simply point to the rear of the
1536                  * buffer, which lead to buffer overlap when doing encryption.
1537                  * usually it's ok and it indeed passed all existing tests.
1538                  * but not sure if there will be subtle problems in the future.
1539                  * so right now we choose to alloc another new buffer. we'll
1540                  * see how it works.
1541                  */
1542 #if 0
1543                 rs->rs_msg = (struct lustre_msg *)
1544                              (rs->rs_repbuf + rs->rs_repbuf_len -
1545                               msgsize - GSS_PRIVBUF_SUFFIX_LEN);
1546 #endif
1547                 char *msgbuf;
1548
1549                 msgsize += GSS_PRIVBUF_PREFIX_LEN + GSS_PRIVBUF_SUFFIX_LEN;
1550                 OBD_ALLOC(msgbuf, msgsize);
1551                 if (!msgbuf) {
1552                         CERROR("can't alloc %d\n", msgsize);
1553                         svcsec_free_reply_state(rs);
1554                         req->rq_reply_state = NULL;
1555                         RETURN(-ENOMEM);
1556                 }
1557                 rs->rs_msg = (struct lustre_msg *)
1558                                 (msgbuf + GSS_PRIVBUF_PREFIX_LEN);
1559         }
1560
1561         req->rq_repmsg = rs->rs_msg;
1562
1563         RETURN(0);
1564 }
1565
1566 static
1567 void gss_svcsec_free_repbuf(struct ptlrpc_svcsec *svcsec,
1568                             struct ptlrpc_reply_state *rs)
1569 {
1570         unsigned long p1 = (unsigned long) rs->rs_msg;
1571         unsigned long p2 = (unsigned long) rs->rs_buf;
1572
1573         LASSERT(rs->rs_buf);
1574         LASSERT(rs->rs_msg);
1575         LASSERT(rs->rs_msg_len);
1576
1577         if (p1 < p2 || p1 >= p2 + rs->rs_buf_len) {
1578                 char *start = (char*) rs->rs_msg - GSS_PRIVBUF_PREFIX_LEN;
1579                 int size = rs->rs_msg_len + GSS_PRIVBUF_PREFIX_LEN +
1580                            GSS_PRIVBUF_SUFFIX_LEN;
1581                 OBD_FREE(start, size);
1582         }
1583
1584         svcsec_free_reply_state(rs);
1585 }
1586
1587 struct ptlrpc_svcsec svcsec_gss = {
1588         .pss_owner              = THIS_MODULE,
1589         .pss_name               = "svcsec.gss",
1590         .pss_flavor             = PTLRPCS_FLVR_MAJOR_GSS,
1591         .accept                 = gss_svcsec_accept,
1592         .authorize              = gss_svcsec_authorize,
1593         .alloc_repbuf           = gss_svcsec_alloc_repbuf,
1594         .free_repbuf            = gss_svcsec_free_repbuf,
1595         .cleanup_req            = gss_svcsec_cleanup_req,
1596 };
1597
1598 /* XXX hacking */
1599 void lgss_svc_cache_purge_all(void)
1600 {
1601         cache_purge(&rsi_cache);
1602         cache_purge(&rsc_cache);
1603 }
1604 EXPORT_SYMBOL(lgss_svc_cache_purge_all);
1605
1606 void lgss_svc_cache_flush(__u32 uid)
1607 {
1608         rsc_flush(uid);
1609 }
1610 EXPORT_SYMBOL(lgss_svc_cache_flush);
1611
1612 int gss_svc_init(void)
1613 {
1614         int rc;
1615
1616         rc = svcsec_register(&svcsec_gss);
1617         if (!rc) {
1618                 cache_register(&rsc_cache);
1619                 cache_register(&rsi_cache);
1620         }
1621         return rc;
1622 }
1623
1624 void gss_svc_exit(void)
1625 {
1626         int rc;
1627         if ((rc = cache_unregister(&rsi_cache)))
1628                 CERROR("unregister rsi cache: %d\n", rc);
1629         if ((rc = cache_unregister(&rsc_cache)))
1630                 CERROR("unregister rsc cache: %d\n", rc);
1631         if ((rc = svcsec_unregister(&svcsec_gss)))
1632                 CERROR("unregister svcsec_gss: %d\n", rc);
1633 }