lustre/sec/gss/sec_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  * linux/net/sunrpc/auth_gss.c
  12  *
  13  * RPCSEC_GSS client authentication.
  14  *
  15  *  Copyright (c) 2000 The Regents of the University of Michigan.
  16  *  All rights reserved.
  17  *
  18  *  Dug Song       <dugsong@monkey.org>
  19  *  Andy Adamson   <andros@umich.edu>
  20  *
  21  *  Redistribution and use in source and binary forms, with or without
  22  *  modification, are permitted provided that the following conditions
  23  *  are met:
  24  *
  25  *  1. Redistributions of source code must retain the above copyright
  26  *     notice, this list of conditions and the following disclaimer.
  27  *  2. Redistributions in binary form must reproduce the above copyright
  28  *     notice, this list of conditions and the following disclaimer in the
  29  *     documentation and/or other materials provided with the distribution.
  30  *  3. Neither the name of the University nor the names of its
  31  *     contributors may be used to endorse or promote products derived
  32  *     from this software without specific prior written permission.
  33  *
  34  *  THIS SOFTWARE IS PROVIDED ``AS IS'' AND ANY EXPRESS OR IMPLIED
  35  *  WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES OF
  36  *  MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
  37  *  DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
  38  *  FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
  39  *  CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
  40  *  SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR
  41  *  BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF
  42  *  LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING
  43  *  NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
  44  *  SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
  45  *
  46  */
  47
  48 #ifndef EXPORT_SYMTAB
  49 # define EXPORT_SYMTAB
  50 #endif
  51 #define DEBUG_SUBSYSTEM S_SEC
  52 #ifdef __KERNEL__
  53 #include <linux/init.h>
  54 #include <linux/module.h>
  55 #include <linux/slab.h>
  56 #include <linux/dcache.h>
  57 #include <linux/fs.h>
  58 #include <linux/random.h>
  59 /* for rpc_pipefs */
  60 struct rpc_clnt;
  61 #include <linux/sunrpc/rpc_pipe_fs.h>
  62 #else
  63 #include <liblustre.h>
  64 #endif
  65
  66 #include <libcfs/kp30.h>
  67 #include <linux/obd.h>
  68 #include <linux/obd_class.h>
  69 #include <linux/obd_support.h>
  70 #include <linux/lustre_idl.h>
  71 #include <linux/lustre_net.h>
  72 #include <linux/lustre_import.h>
  73 #include <linux/lustre_sec.h>
  74
  75 #include "gss_err.h"
  76 #include "gss_internal.h"
  77 #include "gss_api.h"
  78
  79 #define LUSTRE_PIPEDIR          "/lustre"
  80
  81 #define GSS_CREDCACHE_EXPIRE    (30 * 60)          /* 30 minute */
  82
  83 #define GSS_TIMEOUT_DELTA       (5)
  84 #define CRED_REFRESH_UPCALL_TIMEOUT                             \
  85         ({                                                      \
  86                 int timeout = obd_timeout - GSS_TIMEOUT_DELTA;  \
  87                                                                 \
  88                 if (timeout < GSS_TIMEOUT_DELTA * 2)            \
  89                         timeout = GSS_TIMEOUT_DELTA * 2;        \
  90                 timeout;                                        \
  91         })
  92 #define SECINIT_RPC_TIMEOUT                                     \
  93         ({                                                      \
  94                 int timeout = CRED_REFRESH_UPCALL_TIMEOUT -     \
  95                               GSS_TIMEOUT_DELTA;                \
  96                 if (timeout < GSS_TIMEOUT_DELTA)                \
  97                         timeout = GSS_TIMEOUT_DELTA;            \
  98                 timeout;                                        \
  99         })
 100 #define SECFINI_RPC_TIMEOUT     (GSS_TIMEOUT_DELTA)
 101
 102
 103 /**********************************************
 104  * gss security init/fini helper              *
 105  **********************************************/
 106
 107 static int secinit_compose_request(struct obd_import *imp,
 108                                    char *buf, int bufsize,
 109                                    int lustre_srv,
 110                                    uid_t uid, gid_t gid,
 111                                    long token_size,
 112                                    char __user *token)
 113 {
 114         struct ptlrpcs_wire_hdr *hdr;
 115         struct lustre_msg       *lmsg;
 116         struct mds_req_sec_desc *secdesc;
 117         int                      size = sizeof(*secdesc);
 118         __u32                    lmsg_size, *p;
 119         int                      rc;
 120
 121         lmsg_size = lustre_msg_size(1, &size);
 122
 123         if (sizeof(*hdr) + lmsg_size + size_round(token_size) > bufsize) {
 124                 CERROR("token size %ld too large\n", token_size);
 125                 return -EINVAL;
 126         }
 127
 128         /* security wire hdr */
 129         hdr = buf_to_sec_hdr(buf);
 130         hdr->flavor  = cpu_to_le32(PTLRPCS_FLVR_GSS_NONE);
 131         hdr->msg_len = cpu_to_le32(lmsg_size);
 132         hdr->sec_len = cpu_to_le32(8 * 4 + token_size);
 133
 134         /* lustre message & secdesc */
 135         lmsg = buf_to_lustre_msg(buf);
 136
 137         lustre_init_msg(lmsg, 1, &size, NULL);
 138         secdesc = lustre_msg_buf(lmsg, 0, size);
 139         secdesc->rsd_uid = secdesc->rsd_fsuid = uid;
 140         secdesc->rsd_gid = secdesc->rsd_fsgid = gid;
 141         secdesc->rsd_cap = secdesc->rsd_ngroups = 0;
 142
 143         lmsg->handle   = imp->imp_remote_handle;
 144         lmsg->type     = PTL_RPC_MSG_REQUEST;
 145         lmsg->opc      = SEC_INIT;
 146         lmsg->flags    = 0;
 147         lmsg->conn_cnt = imp->imp_conn_cnt;
 148
 149         p = (__u32 *) (buf + sizeof(*hdr) + lmsg_size);
 150
 151         /* gss hdr */
 152         *p++ = cpu_to_le32(PTLRPC_SEC_GSS_VERSION);     /* gss version */
 153         *p++ = cpu_to_le32(PTLRPCS_FLVR_KRB5I);         /* subflavor */
 154         *p++ = cpu_to_le32(PTLRPCS_GSS_PROC_INIT);      /* proc */
 155         *p++ = cpu_to_le32(0);                          /* seq */
 156         *p++ = cpu_to_le32(PTLRPCS_GSS_SVC_NONE);       /* service */
 157         *p++ = cpu_to_le32(0);                          /* context handle */
 158
 159         /* plus lustre svc type */
 160         *p++ = cpu_to_le32(lustre_srv);
 161
 162         /* now the token part */
 163         *p++ = cpu_to_le32((__u32) token_size);
 164         LASSERT(((char *)p - buf) + token_size <= bufsize);
 165
 166         rc = copy_from_user(p, token, token_size);
 167         if (rc) {
 168                 CERROR("can't copy token\n");
 169                 return -EFAULT;
 170         }
 171
 172         rc = size_round(((char *)p - buf) + token_size);
 173         return rc;
 174 }
 175
 176 static int secinit_parse_reply(char *repbuf, int replen,
 177                                char __user *outbuf, long outlen)
 178 {
 179         __u32                   *p = (__u32 *)repbuf;
 180         struct ptlrpcs_wire_hdr *hdr = (struct ptlrpcs_wire_hdr *) repbuf;
 181         __u32                    lmsg_len, sec_len, status;
 182         __u32                    major, minor, seq, obj_len, round_len;
 183         __u32                    effective = 0;
 184
 185         if (replen <= (4 + 6) * 4) {
 186                 CERROR("reply size %d too small\n", replen);
 187                 return -EINVAL;
 188         }
 189
 190         hdr->flavor = le32_to_cpu(hdr->flavor);
 191         hdr->msg_len = le32_to_cpu(hdr->msg_len);
 192         hdr->sec_len = le32_to_cpu(hdr->sec_len);
 193
 194         lmsg_len = le32_to_cpu(p[2]);
 195         sec_len = le32_to_cpu(p[3]);
 196
 197         /* sanity checks */
 198         if (hdr->flavor != PTLRPCS_FLVR_GSS_NONE) {
 199                 CERROR("unexpected reply\n");
 200                 return -EINVAL;
 201         }
 202         if (hdr->msg_len % 8 ||
 203             sizeof(*hdr) + hdr->msg_len + hdr->sec_len > replen) {
 204                 CERROR("unexpected reply\n");
 205                 return -EINVAL;
 206         }
 207         if (hdr->sec_len > outlen) {
 208                 CERROR("outbuf too small\n");
 209                 return -EINVAL;
 210         }
 211
 212         p = (__u32 *) buf_to_sec_data(repbuf);
 213         effective = 0;
 214
 215         status = le32_to_cpu(*p++);
 216         major = le32_to_cpu(*p++);
 217         minor = le32_to_cpu(*p++);
 218         seq = le32_to_cpu(*p++);
 219         effective += 4 * 4;
 220
 221         if (copy_to_user(outbuf, &status, 4))
 222                 return -EFAULT;
 223         outbuf += 4;
 224         if (copy_to_user(outbuf, &major, 4))
 225                 return -EFAULT;
 226         outbuf += 4;
 227         if (copy_to_user(outbuf, &minor, 4))
 228                 return -EFAULT;
 229         outbuf += 4;
 230         if (copy_to_user(outbuf, &seq, 4))
 231                 return -EFAULT;
 232         outbuf += 4;
 233
 234         obj_len = le32_to_cpu(*p++);
 235         round_len = (obj_len + 3) & ~ 3;
 236         if (copy_to_user(outbuf, &obj_len, 4))
 237                 return -EFAULT;
 238         outbuf += 4;
 239         if (copy_to_user(outbuf, (char *)p, round_len))
 240                 return -EFAULT;
 241         p += round_len / 4;
 242         outbuf += round_len;
 243         effective += 4 + round_len;
 244
 245         obj_len = le32_to_cpu(*p++);
 246         round_len = (obj_len + 3) & ~ 3;
 247         if (copy_to_user(outbuf, &obj_len, 4))
 248                 return -EFAULT;
 249         outbuf += 4;
 250         if (copy_to_user(outbuf, (char *)p, round_len))
 251                 return -EFAULT;
 252         p += round_len / 4;
 253         outbuf += round_len;
 254         effective += 4 + round_len;
 255
 256         return effective;
 257 }
 258
 259 /* XXX move to where lgssd could see */
 260 struct lgssd_ioctl_param {
 261         int             version;        /* in   */
 262         char           *uuid;           /* in   */
 263         int             lustre_svc;     /* in   */
 264         uid_t           uid;            /* in   */
 265         gid_t           gid;            /* in   */
 266         long            send_token_size;/* in   */
 267         char           *send_token;     /* in   */
 268         long            reply_buf_size; /* in   */
 269         char           *reply_buf;      /* in   */
 270         long            status;         /* out  */
 271         long            reply_length;   /* out  */
 272 };
 273
 274 static int gss_send_secinit_rpc(__user char *buffer, unsigned long count)
 275 {
 276         struct obd_import        *imp;
 277         struct ptlrpc_request    *request = NULL;
 278         struct lgssd_ioctl_param  param;
 279         const int                 reqbuf_size = 1024;
 280         const int                 repbuf_size = 1024;
 281         char                     *reqbuf, *repbuf;
 282         struct obd_device        *obd;
 283         char                      obdname[64];
 284         long                      lsize;
 285         int                       rc, reqlen, replen;
 286
 287         if (count != sizeof(param)) {
 288                 CERROR("ioctl size %lu, expect %d, please check lgssd version\n",
 289                         count, sizeof(param));
 290                 RETURN(-EINVAL);
 291         }
 292         if (copy_from_user(&param, buffer, sizeof(param))) {
 293                 CERROR("failed copy data from lgssd\n");
 294                 RETURN(-EFAULT);
 295         }
 296
 297         if (param.version != GSSD_INTERFACE_VERSION) {
 298                 CERROR("gssd interface version %d (expect %d)\n",
 299                         param.version, GSSD_INTERFACE_VERSION);
 300                 RETURN(-EINVAL);
 301         }
 302
 303         /* take name */
 304         if (strncpy_from_user(obdname, param.uuid,
 305                               sizeof(obdname)) <= 0) {
 306                 CERROR("Invalid obdname pointer\n");
 307                 RETURN(-EFAULT);
 308         }
 309
 310         obd = class_name2obd(obdname);
 311         if (!obd) {
 312                 CERROR("no such obd %s\n", obdname);
 313                 RETURN(-EINVAL);
 314         }
 315
 316         imp = class_import_get(obd->u.cli.cl_import);
 317
 318         OBD_ALLOC(reqbuf, reqbuf_size);
 319         OBD_ALLOC(repbuf, reqbuf_size);
 320
 321         if (!reqbuf || !repbuf) {
 322                 CERROR("Can't alloc buffer: %p/%p\n", reqbuf, repbuf);
 323                 param.status = -ENOMEM;
 324                 goto out_copy;
 325         }
 326
 327         /* get token */
 328         reqlen = secinit_compose_request(imp, reqbuf, reqbuf_size,
 329                                          param.lustre_svc,
 330                                          param.uid, param.gid,
 331                                          param.send_token_size,
 332                                          param.send_token);
 333         if (reqlen < 0) {
 334                 param.status = reqlen;
 335                 goto out_copy;
 336         }
 337
 338         request = ptl_do_rawrpc(imp, reqbuf, reqbuf_size, reqlen,
 339                                 repbuf, repbuf_size, &replen,
 340                                 SECINIT_RPC_TIMEOUT, &rc);
 341         if (request == NULL || rc) {
 342                 param.status = rc;
 343                 goto out_copy;
 344         }
 345
 346         if (replen > param.reply_buf_size) {
 347                 CERROR("output buffer size %ld too small, need %d\n",
 348                         param.reply_buf_size, replen);
 349                 param.status = -EINVAL;
 350                 goto out_copy;
 351         }
 352
 353         lsize = secinit_parse_reply(repbuf, replen,
 354                                     param.reply_buf, param.reply_buf_size);
 355         if (lsize < 0) {
 356                 param.status = (int) lsize;
 357                 goto out_copy;
 358         }
 359
 360         param.status = 0;
 361         param.reply_length = lsize;
 362
 363 out_copy:
 364         if (copy_to_user(buffer, &param, sizeof(param)))
 365                 rc = -EFAULT;
 366         else
 367                 rc = 0;
 368
 369         class_import_put(imp);
 370         if (request == NULL) {
 371                 if (repbuf)
 372                         OBD_FREE(repbuf, repbuf_size);
 373                 if (reqbuf)
 374                         OBD_FREE(reqbuf, reqbuf_size);
 375         } else {
 376                 rawrpc_req_finished(request);
 377         }
 378         RETURN(rc);
 379 }
 380
 381 /**********************************************
 382  * structure definitions                      *
 383  **********************************************/
 384 struct gss_sec {
 385         struct ptlrpc_sec       gs_base;
 386         struct gss_api_mech    *gs_mech;
 387 #ifdef __KERNEL__
 388         spinlock_t              gs_lock;
 389         struct list_head        gs_upcalls;
 390         char                   *gs_pipepath;
 391         struct dentry          *gs_depipe;
 392 #endif
 393 };
 394
 395 #ifdef __KERNEL__
 396
 397 static rwlock_t gss_ctx_lock = RW_LOCK_UNLOCKED;
 398
 399 struct gss_upcall_msg_data {
 400         __u32                           gum_uid;
 401         __u32                           gum_svc;
 402         __u32                           gum_nal;
 403         __u32                           gum_netid;
 404         __u64                           gum_nid;
 405 };
 406
 407 struct gss_upcall_msg {
 408         struct rpc_pipe_msg             gum_base;
 409         atomic_t                        gum_refcount;
 410         struct list_head                gum_list;
 411         struct gss_sec                 *gum_gsec;
 412         wait_queue_head_t               gum_waitq;
 413         char                            gum_obdname[64];
 414         struct gss_upcall_msg_data      gum_data;
 415 };
 416
 417 /**********************************************
 418  * rpc_pipe upcall helpers                    *
 419  **********************************************/
 420 static
 421 void gss_release_msg(struct gss_upcall_msg *gmsg)
 422 {
 423         ENTRY;
 424         LASSERT(atomic_read(&gmsg->gum_refcount) > 0);
 425
 426         if (!atomic_dec_and_test(&gmsg->gum_refcount)) {
 427                 CDEBUG(D_SEC, "gmsg %p ref %d\n", gmsg,
 428                        atomic_read(&gmsg->gum_refcount));
 429                 EXIT;
 430                 return;
 431         }
 432         LASSERT(list_empty(&gmsg->gum_list));
 433 #if 0
 434         LASSERT(list_empty(&gmsg->gum_base.list));
 435 #else
 436         /* XXX */
 437         if (!list_empty(&gmsg->gum_base.list)) {
 438                 CWARN("msg %p: list: %p/%p/%p, copied %d, err %d, wq %d\n",
 439                       gmsg, &gmsg->gum_base.list,
 440                       gmsg->gum_base.list.prev, gmsg->gum_base.list.next,
 441                       gmsg->gum_base.copied, gmsg->gum_base.errno,
 442                       list_empty(&gmsg->gum_waitq.task_list));
 443                 LBUG();
 444         }
 445 #endif
 446         OBD_FREE(gmsg, sizeof(*gmsg));
 447         EXIT;
 448 }
 449
 450 static void
 451 gss_unhash_msg_nolock(struct gss_upcall_msg *gmsg)
 452 {
 453         LASSERT(spin_is_locked(&gmsg->gum_gsec->gs_lock));
 454
 455         if (list_empty(&gmsg->gum_list))
 456                 return;
 457
 458         list_del_init(&gmsg->gum_list);
 459         wake_up(&gmsg->gum_waitq);
 460         LASSERT(atomic_read(&gmsg->gum_refcount) > 1);
 461         atomic_dec(&gmsg->gum_refcount);
 462 }
 463
 464 static void
 465 gss_unhash_msg(struct gss_upcall_msg *gmsg)
 466 {
 467         struct gss_sec *gsec = gmsg->gum_gsec;
 468
 469         spin_lock(&gsec->gs_lock);
 470         gss_unhash_msg_nolock(gmsg);
 471         spin_unlock(&gsec->gs_lock);
 472 }
 473
 474 static
 475 struct gss_upcall_msg * gss_find_upcall(struct gss_sec *gsec,
 476                                         char *obdname,
 477                                         struct gss_upcall_msg_data *gmd)
 478 {
 479         struct gss_upcall_msg *gmsg;
 480         ENTRY;
 481
 482         LASSERT(spin_is_locked(&gsec->gs_lock));
 483
 484         list_for_each_entry(gmsg, &gsec->gs_upcalls, gum_list) {
 485                 if (memcmp(&gmsg->gum_data, gmd, sizeof(*gmd)))
 486                         continue;
 487                 if (strcmp(gmsg->gum_obdname, obdname))
 488                         continue;
 489                 LASSERT(atomic_read(&gmsg->gum_refcount) > 0);
 490                 atomic_inc(&gmsg->gum_refcount);
 491                 CDEBUG(D_SEC, "found gmsg at %p: obdname %s, uid %d, ref %d\n",
 492                        gmsg, obdname, gmd->gum_uid,
 493                        atomic_read(&gmsg->gum_refcount));
 494                 RETURN(gmsg);
 495         }
 496         RETURN(NULL);
 497 }
 498
 499 static void gss_init_upcall_msg(struct gss_upcall_msg *gmsg,
 500                                 struct gss_sec *gsec, char *obdname,
 501                                 struct gss_upcall_msg_data *gmd)
 502 {
 503         struct rpc_pipe_msg *rpcmsg;
 504         ENTRY;
 505
 506         /* 2 refs: 1 for hash, 1 for current user */
 507         init_waitqueue_head(&gmsg->gum_waitq);
 508         list_add(&gmsg->gum_list, &gsec->gs_upcalls);
 509         atomic_set(&gmsg->gum_refcount, 2);
 510         gmsg->gum_gsec = gsec;
 511         strncpy(gmsg->gum_obdname, obdname, sizeof(gmsg->gum_obdname));
 512         memcpy(&gmsg->gum_data, gmd, sizeof(*gmd));
 513
 514         rpcmsg = &gmsg->gum_base;
 515         INIT_LIST_HEAD(&rpcmsg->list);
 516         rpcmsg->data = &gmsg->gum_data;
 517         rpcmsg->len = sizeof(gmsg->gum_data);
 518         rpcmsg->copied = 0;
 519         rpcmsg->errno = 0;
 520         EXIT;
 521 }
 522 #endif /* __KERNEL__ */
 523
 524 /********************************************
 525  * gss cred manipulation helpers            *
 526  ********************************************/
 527 #if 0
 528 static
 529 int gss_cred_is_uptodate_ctx(struct ptlrpc_cred *cred)
 530 {
 531         struct gss_cred *gcred = container_of(cred, struct gss_cred, gc_base);
 532         int res = 0;
 533
 534         read_lock(&gss_ctx_lock);
 535         if (((cred->pc_flags & PTLRPC_CRED_FLAGS_MASK) ==
 536              PTLRPC_CRED_UPTODATE) &&
 537             gcred->gc_ctx)
 538                 res = 1;
 539         read_unlock(&gss_ctx_lock);
 540         return res;
 541 }
 542 #endif
 543
 544 static inline
 545 struct gss_cl_ctx * gss_get_ctx(struct gss_cl_ctx *ctx)
 546 {
 547         atomic_inc(&ctx->gc_refcount);
 548         return ctx;
 549 }
 550
 551 static
 552 void gss_destroy_ctx(struct gss_cl_ctx *ctx)
 553 {
 554         ENTRY;
 555
 556         CDEBUG(D_SEC, "destroy cl_ctx %p\n", ctx);
 557         if (ctx->gc_gss_ctx)
 558                 kgss_delete_sec_context(&ctx->gc_gss_ctx);
 559
 560         if (ctx->gc_wire_ctx.len > 0) {
 561                 OBD_FREE(ctx->gc_wire_ctx.data, ctx->gc_wire_ctx.len);
 562                 ctx->gc_wire_ctx.len = 0;
 563         }
 564
 565         OBD_FREE(ctx, sizeof(*ctx));
 566 }
 567
 568 static
 569 void gss_put_ctx(struct gss_cl_ctx *ctx)
 570 {
 571         if (atomic_dec_and_test(&ctx->gc_refcount))
 572                 gss_destroy_ctx(ctx);
 573 }
 574
 575 static
 576 struct gss_cl_ctx *gss_cred_get_ctx(struct ptlrpc_cred *cred)
 577 {
 578         struct gss_cred *gcred = container_of(cred, struct gss_cred, gc_base);
 579         struct gss_cl_ctx *ctx = NULL;
 580
 581         read_lock(&gss_ctx_lock);
 582         if (gcred->gc_ctx)
 583                 ctx = gss_get_ctx(gcred->gc_ctx);
 584         read_unlock(&gss_ctx_lock);
 585         return ctx;
 586 }
 587
 588 static
 589 void gss_cred_set_ctx(struct ptlrpc_cred *cred, struct gss_cl_ctx *ctx)
 590 {
 591         struct gss_cred *gcred = container_of(cred, struct gss_cred, gc_base);
 592         struct gss_cl_ctx *old;
 593         __u64 ctx_expiry;
 594         ENTRY;
 595
 596         if (kgss_inquire_context(ctx->gc_gss_ctx, &ctx_expiry)) {
 597                 CERROR("unable to get expire time\n");
 598                 ctx_expiry = 1; /* make it expired now */
 599         }
 600         cred->pc_expire = (unsigned long) ctx_expiry;
 601
 602         write_lock(&gss_ctx_lock);
 603         old = gcred->gc_ctx;
 604         gcred->gc_ctx = ctx;
 605         set_bit(PTLRPC_CRED_UPTODATE_BIT, &cred->pc_flags);
 606         write_unlock(&gss_ctx_lock);
 607         if (old)
 608                 gss_put_ctx(old);
 609
 610         CDEBUG(D_SEC, "client refreshed gss cred %p(uid %u)\n",
 611                cred, cred->pc_uid);
 612         EXIT;
 613 }
 614
 615 static int
 616 simple_get_bytes(char **buf, __u32 *buflen, void *res, __u32 reslen)
 617 {
 618         if (*buflen < reslen) {
 619                 CERROR("buflen %u < %u\n", *buflen, reslen);
 620                 return -EINVAL;
 621         }
 622
 623         memcpy(res, *buf, reslen);
 624         *buf += reslen;
 625         *buflen -= reslen;
 626         return 0;
 627 }
 628
 629 /* data passed down:
 630  *  - uid
 631  *  - timeout
 632  *  - gc_win / error
 633  *  - wire_ctx (rawobj)
 634  *  - mech_ctx? (rawobj)
 635  */
 636 static
 637 int gss_parse_init_downcall(struct gss_api_mech *gm, rawobj_t *buf,
 638                             struct gss_cl_ctx **gc,
 639                             struct gss_upcall_msg_data *gmd, int *gss_err)
 640 {
 641         char *p = (char *)buf->data;
 642         struct gss_cl_ctx *ctx;
 643         __u32 len = buf->len;
 644         unsigned int timeout;
 645         rawobj_t tmp_buf;
 646         int err = -EPERM;
 647         ENTRY;
 648
 649         *gc = NULL;
 650         *gss_err = 0;
 651
 652         OBD_ALLOC(ctx, sizeof(*ctx));
 653         if (!ctx)
 654                 RETURN(-ENOMEM);
 655
 656         ctx->gc_proc = RPC_GSS_PROC_DATA;
 657         ctx->gc_seq = 0;
 658         spin_lock_init(&ctx->gc_seq_lock);
 659         atomic_set(&ctx->gc_refcount,1);
 660
 661         if (simple_get_bytes(&p, &len, &gmd->gum_uid, sizeof(gmd->gum_uid)))
 662                 goto err_free_ctx;
 663         if (simple_get_bytes(&p, &len, &gmd->gum_svc, sizeof(gmd->gum_svc)))
 664                 goto err_free_ctx;
 665         if (simple_get_bytes(&p, &len, &gmd->gum_nal, sizeof(gmd->gum_nal)))
 666                 goto err_free_ctx;
 667         if (simple_get_bytes(&p, &len, &gmd->gum_netid, sizeof(gmd->gum_netid)))
 668                 goto err_free_ctx;
 669         if (simple_get_bytes(&p, &len, &gmd->gum_nid, sizeof(gmd->gum_nid)))
 670                 goto err_free_ctx;
 671         /* FIXME: discarded timeout for now */
 672         if (simple_get_bytes(&p, &len, &timeout, sizeof(timeout)))
 673                 goto err_free_ctx;
 674         if (simple_get_bytes(&p, &len, &ctx->gc_win, sizeof(ctx->gc_win)))
 675                 goto err_free_ctx;
 676
 677         /* lgssd signals an error by passing ctx->gc_win = 0: */
 678         if (!ctx->gc_win) {
 679                 /* in which case the next 2 int are:
 680                  * - rpc error
 681                  * - gss error
 682                  */
 683                 if (simple_get_bytes(&p, &len, &err, sizeof(err))) {
 684                         err = -EPERM;
 685                         goto err_free_ctx;
 686                 }
 687                 if (simple_get_bytes(&p, &len, gss_err, sizeof(*gss_err))) {
 688                         err = -EPERM;
 689                         goto err_free_ctx;
 690                 }
 691                 if (err == 0 && *gss_err == 0) {
 692                         CERROR("no error passed from downcall\n");
 693                         err = -EPERM;
 694                 }
 695                 goto err_free_ctx;
 696         }
 697
 698         if (rawobj_extract_local(&tmp_buf, (__u32 **) ((void *)&p), &len))
 699                 goto err_free_ctx;
 700         if (rawobj_dup(&ctx->gc_wire_ctx, &tmp_buf)) {
 701                 err = -ENOMEM;
 702                 goto err_free_ctx;
 703         }
 704         if (rawobj_extract_local(&tmp_buf, (__u32 **) ((void *)&p), &len))
 705                 goto err_free_wire_ctx;
 706         if (len) {
 707                 CERROR("unexpected trailing %u bytes\n", len);
 708                 goto err_free_wire_ctx;
 709         }
 710         if (kgss_import_sec_context(&tmp_buf, gm, &ctx->gc_gss_ctx))
 711                 goto err_free_wire_ctx;
 712
 713         *gc = ctx;
 714         RETURN(0);
 715
 716 err_free_wire_ctx:
 717         if (ctx->gc_wire_ctx.data)
 718                 OBD_FREE(ctx->gc_wire_ctx.data, ctx->gc_wire_ctx.len);
 719 err_free_ctx:
 720         OBD_FREE(ctx, sizeof(*ctx));
 721         CDEBUG(D_SEC, "err_code %d, gss code %d\n", err, *gss_err);
 722         return err;
 723 }
 724
 725 /***************************************
 726  * cred APIs                           *
 727  ***************************************/
 728 #ifdef __KERNEL__
 729 static int gss_cred_refresh(struct ptlrpc_cred *cred)
 730 {
 731         struct obd_import          *import;
 732         struct gss_sec             *gsec;
 733         struct gss_upcall_msg      *gss_msg, *gss_new;
 734         struct gss_upcall_msg_data  gmd;
 735         struct dentry              *dentry;
 736         char                       *obdname, *obdtype;
 737         wait_queue_t                wait;
 738         uid_t                       uid = cred->pc_uid;
 739         int                         res;
 740         ENTRY;
 741
 742         might_sleep();
 743
 744         /* any flags means it has been handled, do nothing */
 745         if (cred->pc_flags & PTLRPC_CRED_FLAGS_MASK)
 746                 RETURN(0);
 747
 748         LASSERT(cred->pc_sec);
 749         LASSERT(cred->pc_sec->ps_import);
 750         LASSERT(cred->pc_sec->ps_import->imp_obd);
 751
 752         import = cred->pc_sec->ps_import;
 753         if (!import->imp_connection) {
 754                 CERROR("import has no connection set\n");
 755                 RETURN(-EINVAL);
 756         }
 757
 758         gmd.gum_uid = uid;
 759         gmd.gum_nal = import->imp_connection->c_peer.peer_ni->pni_number;
 760         gmd.gum_netid = 0;
 761         gmd.gum_nid = import->imp_connection->c_peer.peer_id.nid;
 762
 763         obdtype = import->imp_obd->obd_type->typ_name;
 764         if (!strcmp(obdtype, "mdc"))
 765                 gmd.gum_svc = LUSTRE_GSS_SVC_MDS;
 766         else if (!strcmp(obdtype, "osc"))
 767                 gmd.gum_svc = LUSTRE_GSS_SVC_OSS;
 768         else {
 769                 CERROR("gss on %s?\n", obdtype);
 770                 RETURN(-EINVAL);
 771         }
 772
 773         gsec = container_of(cred->pc_sec, struct gss_sec, gs_base);
 774         obdname = import->imp_obd->obd_name;
 775         dentry = gsec->gs_depipe;
 776         gss_new = NULL;
 777         res = 0;
 778
 779         CDEBUG(D_SEC, "Initiate gss context %p(%u@%s)\n",
 780                container_of(cred, struct gss_cred, gc_base),
 781                uid, import->imp_target_uuid.uuid);
 782
 783 again:
 784         spin_lock(&gsec->gs_lock);
 785         gss_msg = gss_find_upcall(gsec, obdname, &gmd);
 786         if (gss_msg) {
 787                 if (gss_new) {
 788                         OBD_FREE(gss_new, sizeof(*gss_new));
 789                         gss_new = NULL;
 790                 }
 791                 GOTO(waiting, res);
 792         }
 793
 794         if (!gss_new) {
 795                 spin_unlock(&gsec->gs_lock);
 796                 OBD_ALLOC(gss_new, sizeof(*gss_new));
 797                 if (!gss_new)
 798                         RETURN(-ENOMEM);
 799                 goto again;
 800         }
 801         /* so far we'v created gss_new */
 802         gss_init_upcall_msg(gss_new, gsec, obdname, &gmd);
 803
 804         /* we'v created upcall msg, nobody else should touch the
 805          * flag of this cred, unless be set as dead/expire by
 806          * administrator via lctl etc.
 807          */
 808         if (cred->pc_flags & PTLRPC_CRED_FLAGS_MASK) {
 809                 CWARN("cred %p("LPU64"/%u) was set flags %lx unexpectedly\n",
 810                       cred, cred->pc_pag, cred->pc_uid, cred->pc_flags);
 811                 cred->pc_flags |= PTLRPC_CRED_DEAD | PTLRPC_CRED_ERROR;
 812                 gss_unhash_msg_nolock(gss_new);
 813                 spin_unlock(&gsec->gs_lock);
 814                 gss_release_msg(gss_new);
 815                 RETURN(0);
 816         }
 817
 818         /* need to make upcall now */
 819         spin_unlock(&gsec->gs_lock);
 820         res = rpc_queue_upcall(dentry->d_inode, &gss_new->gum_base);
 821         if (res) {
 822                 CERROR("rpc_queue_upcall failed: %d\n", res);
 823                 gss_unhash_msg(gss_new);
 824                 gss_release_msg(gss_new);
 825                 cred->pc_flags |= PTLRPC_CRED_DEAD | PTLRPC_CRED_ERROR;
 826                 RETURN(res);
 827         }
 828         gss_msg = gss_new;
 829         spin_lock(&gsec->gs_lock);
 830
 831 waiting:
 832         /* upcall might finish quickly */
 833         if (list_empty(&gss_msg->gum_list)) {
 834                 spin_unlock(&gsec->gs_lock);
 835                 res = 0;
 836                 goto out;
 837         }
 838
 839         init_waitqueue_entry(&wait, current);
 840         set_current_state(TASK_INTERRUPTIBLE);
 841         add_wait_queue(&gss_msg->gum_waitq, &wait);
 842         spin_unlock(&gsec->gs_lock);
 843
 844         if (gss_new)
 845                 res = schedule_timeout(CRED_REFRESH_UPCALL_TIMEOUT * HZ);
 846         else {
 847                 schedule();
 848                 res = 0;
 849         }
 850
 851         remove_wait_queue(&gss_msg->gum_waitq, &wait);
 852
 853         /* - the one who refresh the cred for us should also be responsible
 854          *   to set the status of cred, we can simply return.
 855          * - if cred flags has been set, we also don't need to do that again,
 856          *   no matter signal pending or timeout etc.
 857          */
 858         if (!gss_new || cred->pc_flags & PTLRPC_CRED_FLAGS_MASK)
 859                 goto out;
 860
 861         if (signal_pending(current)) {
 862                 CERROR("%s: cred %p: interrupted upcall\n",
 863                        current->comm, cred);
 864                 cred->pc_flags |= PTLRPC_CRED_DEAD | PTLRPC_CRED_ERROR;
 865                 res = -EINTR;
 866         } else if (res == 0) {
 867                 CERROR("cred %p: upcall timedout\n", cred);
 868                 set_bit(PTLRPC_CRED_DEAD_BIT, &cred->pc_flags);
 869                 res = -ETIMEDOUT;
 870         } else
 871                 res = 0;
 872
 873 out:
 874         gss_release_msg(gss_msg);
 875
 876         RETURN(res);
 877 }
 878 #else /* !__KERNEL__ */
 879 extern int lgss_handle_krb5_upcall(uid_t uid, __u32 dest_ip,
 880                                    char *obd_name,
 881                                    char *buf, int bufsize,
 882                                    int (*callback)(char*, unsigned long));
 883
 884 static int gss_cred_refresh(struct ptlrpc_cred *cred)
 885 {
 886         char                    buf[4096];
 887         rawobj_t                obj;
 888         struct obd_import      *imp;
 889         struct gss_sec         *gsec;
 890         struct gss_api_mech    *mech;
 891         struct gss_cl_ctx      *ctx = NULL;
 892         struct vfs_cred         vcred = { 0 };
 893         ptl_nid_t               peer_nid;
 894         __u32                   dest_ip;
 895         __u32                   subflavor;
 896         int                     rc, gss_err;
 897
 898         LASSERT(cred);
 899         LASSERT(cred->pc_sec);
 900         LASSERT(cred->pc_sec->ps_import);
 901         LASSERT(cred->pc_sec->ps_import->imp_obd);
 902
 903         if (ptlrpcs_cred_is_uptodate(cred))
 904                 RETURN(0);
 905
 906         imp = cred->pc_sec->ps_import;
 907         peer_nid = imp->imp_connection->c_peer.peer_id.nid;
 908         dest_ip = (__u32) (peer_nid & 0xFFFFFFFF);
 909         subflavor = cred->pc_sec->ps_flavor.subflavor;
 910
 911         if (subflavor != PTLRPC_SEC_GSS_KRB5I) {
 912                 CERROR("unknown subflavor %u\n", subflavor);
 913                 GOTO(err_out, rc = -EINVAL);
 914         }
 915
 916         rc = lgss_handle_krb5_upcall(cred->pc_uid, dest_ip,
 917                                      imp->imp_obd->obd_name,
 918                                      buf, sizeof(buf),
 919                                      gss_send_secinit_rpc);
 920         LASSERT(rc != 0);
 921         if (rc < 0)
 922                 goto err_out;
 923
 924         obj.data = buf;
 925         obj.len = rc;
 926
 927         gsec = container_of(cred->pc_sec, struct gss_sec, gs_base);
 928         mech = gsec->gs_mech;
 929         LASSERT(mech);
 930         rc = gss_parse_init_downcall(mech, &obj, &ctx, &vcred, &dest_ip,
 931                                      &gss_err);
 932         if (rc || gss_err) {
 933                 CERROR("parse init downcall: rpc %d, gss 0x%x\n", rc, gss_err);
 934                 if (rc != -ERESTART || gss_err != 0)
 935                         set_bit(PTLRPC_CRED_ERROR_BIT, &cred->pc_flags);
 936                 if (rc == 0)
 937                         rc = -EPERM;
 938                 goto err_out;
 939         }
 940
 941         LASSERT(ctx);
 942         gss_cred_set_ctx(cred, ctx);
 943         LASSERT(gss_cred_is_uptodate_ctx(cred));
 944
 945         return 0;
 946 err_out:
 947         set_bit(PTLRPC_CRED_DEAD_BIT, &cred->pc_flags);
 948         return rc;
 949 }
 950 #endif
 951
 952 static int gss_cred_match(struct ptlrpc_cred *cred,
 953                           struct vfs_cred *vcred)
 954 {
 955         RETURN(cred->pc_pag == vcred->vc_pag);
 956 }
 957
 958 static int gss_cred_sign(struct ptlrpc_cred *cred,
 959                          struct ptlrpc_request *req)
 960 {
 961         struct gss_cred         *gcred;
 962         struct gss_cl_ctx       *ctx;
 963         rawobj_t                 lmsg, mic;
 964         __u32                   *vp, *vpsave, vlen, seclen;
 965         __u32                    seqnum, major, rc = 0;
 966         ENTRY;
 967
 968         LASSERT(req->rq_reqbuf);
 969         LASSERT(req->rq_cred == cred);
 970
 971         gcred = container_of(cred, struct gss_cred, gc_base);
 972         ctx = gss_cred_get_ctx(cred);
 973         if (!ctx) {
 974                 CERROR("cred %p("LPU64"/%u) invalidated?\n",
 975                         cred, cred->pc_pag, cred->pc_uid);
 976                 RETURN(-EPERM);
 977         }
 978
 979         lmsg.len = req->rq_reqlen;
 980         lmsg.data = (__u8 *) req->rq_reqmsg;
 981
 982         vp = (__u32 *) (lmsg.data + lmsg.len);
 983         vlen = req->rq_reqbuf_len - sizeof(struct ptlrpcs_wire_hdr) -
 984                lmsg.len;
 985         seclen = vlen;
 986
 987         if (vlen < 6 * 4 + size_round4(ctx->gc_wire_ctx.len)) {
 988                 CERROR("vlen %d, need %d\n",
 989                         vlen, 6 * 4 + size_round4(ctx->gc_wire_ctx.len));
 990                 rc = -EIO;
 991                 goto out;
 992         }
 993
 994         spin_lock(&ctx->gc_seq_lock);
 995         seqnum = ctx->gc_seq++;
 996         spin_unlock(&ctx->gc_seq_lock);
 997
 998         *vp++ = cpu_to_le32(PTLRPC_SEC_GSS_VERSION);    /* version */
 999         *vp++ = cpu_to_le32(PTLRPCS_FLVR_KRB5I);        /* subflavor */
1000         *vp++ = cpu_to_le32(ctx->gc_proc);              /* proc */
1001         *vp++ = cpu_to_le32(seqnum);                    /* seq */
1002         *vp++ = cpu_to_le32(PTLRPCS_GSS_SVC_INTEGRITY); /* service */
1003         vlen -= 5 * 4;
1004
1005         if (rawobj_serialize(&ctx->gc_wire_ctx, &vp, &vlen)) {
1006                 rc = -EIO;
1007                 goto out;
1008         }
1009         CDEBUG(D_SEC, "encoded wire_ctx length %d\n", ctx->gc_wire_ctx.len);
1010
1011         vpsave = vp++;  /* reserve for size */
1012         vlen -= 4;
1013
1014         mic.len = vlen;
1015         mic.data = (unsigned char *)vp;
1016
1017         CDEBUG(D_SEC, "reqbuf at %p, lmsg at %p, len %d, mic at %p, len %d\n",
1018                req->rq_reqbuf, lmsg.data, lmsg.len, mic.data, mic.len);
1019         major = kgss_get_mic(ctx->gc_gss_ctx, GSS_C_QOP_DEFAULT, &lmsg, &mic);
1020         if (major) {
1021                 CERROR("cred %p: req %p compute mic error, major %x\n",
1022                        cred, req, major);
1023                 rc = -EACCES;
1024                 goto out;
1025         }
1026
1027         *vpsave = cpu_to_le32(mic.len);
1028
1029         seclen = seclen - vlen + mic.len;
1030         buf_to_sec_hdr(req->rq_reqbuf)->sec_len = cpu_to_le32(seclen);
1031         req->rq_reqdata_len += size_round(seclen);
1032         CDEBUG(D_SEC, "msg size %d, checksum size %d, total sec size %d\n",
1033                lmsg.len, mic.len, seclen);
1034 out:
1035         gss_put_ctx(ctx);
1036         RETURN(rc);
1037 }
1038
1039 static int gss_cred_verify(struct ptlrpc_cred *cred,
1040                            struct ptlrpc_request *req)
1041 {
1042         struct gss_cred        *gcred;
1043         struct gss_cl_ctx      *ctx;
1044         struct ptlrpcs_wire_hdr *sec_hdr;
1045         rawobj_t                lmsg, mic;
1046         __u32                   *vp, vlen, subflavor, proc, seq, svc;
1047         __u32                   major, minor, rc;
1048         ENTRY;
1049
1050         LASSERT(req->rq_repbuf);
1051         LASSERT(req->rq_cred == cred);
1052
1053         sec_hdr = buf_to_sec_hdr(req->rq_repbuf);
1054         vp = (__u32 *) (req->rq_repbuf + sizeof(*sec_hdr) + sec_hdr->msg_len);
1055         vlen = sec_hdr->sec_len;
1056
1057         if (vlen < 7 * 4) {
1058                 CERROR("reply sec size %u too small\n", vlen);
1059                 RETURN(-EPROTO);
1060         }
1061
1062         if (*vp++ != cpu_to_le32(PTLRPC_SEC_GSS_VERSION)) {
1063                 CERROR("reply have different gss version\n");
1064                 RETURN(-EPROTO);
1065         }
1066         subflavor = le32_to_cpu(*vp++);
1067         proc = le32_to_cpu(*vp++);
1068         vlen -= 3 * 4;
1069
1070         switch (proc) {
1071         case PTLRPCS_GSS_PROC_DATA:
1072                 seq = le32_to_cpu(*vp++);
1073                 svc = le32_to_cpu(*vp++);
1074                 if (svc != PTLRPCS_GSS_SVC_INTEGRITY) {
1075                         CERROR("Unknown svc %d\n", svc);
1076                         RETURN(-EPROTO);
1077                 }
1078                 if (*vp++ != 0) {
1079                         CERROR("Unexpected ctx handle\n");
1080                         RETURN(-EPROTO);
1081                 }
1082                 mic.len = le32_to_cpu(*vp++);
1083                 vlen -= 4 * 4;
1084                 if (vlen < mic.len) {
1085                         CERROR("vlen %d, mic.len %d\n", vlen, mic.len);
1086                         RETURN(-EINVAL);
1087                 }
1088                 mic.data = (unsigned char *)vp;
1089
1090                 gcred = container_of(cred, struct gss_cred, gc_base);
1091                 ctx = gss_cred_get_ctx(cred);
1092                 LASSERT(ctx);
1093
1094                 lmsg.len = sec_hdr->msg_len;
1095                 lmsg.data = (__u8 *) buf_to_lustre_msg(req->rq_repbuf);
1096
1097                 major = kgss_verify_mic(ctx->gc_gss_ctx, &lmsg, &mic, NULL);
1098                 if (major != GSS_S_COMPLETE) {
1099                         CERROR("cred %p: req %p verify mic error: major %x\n",
1100                                cred, req, major);
1101
1102                         if (major == GSS_S_CREDENTIALS_EXPIRED ||
1103                             major == GSS_S_CONTEXT_EXPIRED) {
1104                                 ptlrpcs_cred_expire(cred);
1105                                 req->rq_ptlrpcs_restart = 1;
1106                                 rc = 0;
1107                         } else
1108                                 rc = -EINVAL;
1109
1110                         GOTO(proc_data_out, rc);
1111                 }
1112
1113                 req->rq_repmsg = (struct lustre_msg *) lmsg.data;
1114                 req->rq_replen = lmsg.len;
1115
1116                 /* here we could check the seq number is the same one
1117                  * we sent to server. but portals has prevent us from
1118                  * replay attack, so maybe we don't need check it again.
1119                  */
1120                 rc = 0;
1121 proc_data_out:
1122                 gss_put_ctx(ctx);
1123                 break;
1124         case PTLRPCS_GSS_PROC_ERR:
1125                 major = le32_to_cpu(*vp++);
1126                 minor = le32_to_cpu(*vp++);
1127                 /* server return NO_CONTEXT might be caused by context expire
1128                  * or server reboot/failover. we refresh the cred transparently
1129                  * to upper layer.
1130                  * In some cases, our gss handle is possible to be incidentally
1131                  * identical to another handle since the handle itself is not
1132                  * fully random. In krb5 case, the GSS_S_BAD_SIG will be
1133                  * returned, maybe other gss error for other mechanism. Here we
1134                  * only consider krb5 mech (FIXME) and try to establish new
1135                  * context.
1136                  */
1137                 if (major == GSS_S_NO_CONTEXT ||
1138                     major == GSS_S_BAD_SIG) {
1139                         CWARN("req %p: server report cred %p %s, expired?\n",
1140                                req, cred, (major == GSS_S_NO_CONTEXT) ?
1141                                            "NO_CONTEXT" : "BAD_SIG");
1142
1143                         ptlrpcs_cred_expire(cred);
1144                         req->rq_ptlrpcs_restart = 1;
1145                         rc = 0;
1146                 } else {
1147                         CERROR("req %p: unrecognized gss error (%x/%x)\n",
1148                                 req, major, minor);
1149                         rc = -EACCES;
1150                 }
1151                 break;
1152         default:
1153                 CERROR("unknown gss proc %d\n", proc);
1154                 rc = -EPROTO;
1155         }
1156
1157         RETURN(rc);
1158 }
1159
1160 static int gss_cred_seal(struct ptlrpc_cred *cred,
1161                          struct ptlrpc_request *req)
1162 {
1163         struct gss_cred         *gcred;
1164         struct gss_cl_ctx       *ctx;
1165         struct ptlrpcs_wire_hdr *sec_hdr;
1166         rawobj_buf_t             msg_buf;
1167         rawobj_t                 cipher_buf;
1168         __u32                   *vp, *vpsave, vlen, seclen;
1169         __u32                    major, seqnum, rc = 0;
1170         ENTRY;
1171
1172         LASSERT(req->rq_reqbuf);
1173         LASSERT(req->rq_cred == cred);
1174
1175         gcred = container_of(cred, struct gss_cred, gc_base);
1176         ctx = gss_cred_get_ctx(cred);
1177         if (!ctx) {
1178                 CERROR("cred %p("LPU64"/%u) invalidated?\n",
1179                         cred, cred->pc_pag, cred->pc_uid);
1180                 RETURN(-EPERM);
1181         }
1182
1183         vp = (__u32 *) (req->rq_reqbuf + sizeof(*sec_hdr));
1184         vlen = req->rq_reqbuf_len - sizeof(*sec_hdr);
1185         seclen = vlen;
1186
1187         if (vlen < 6 * 4 + size_round4(ctx->gc_wire_ctx.len)) {
1188                 CERROR("vlen %d, need %d\n",
1189                         vlen, 6 * 4 + size_round4(ctx->gc_wire_ctx.len));
1190                 rc = -EIO;
1191                 goto out;
1192         }
1193
1194         spin_lock(&ctx->gc_seq_lock);
1195         seqnum = ctx->gc_seq++;
1196         spin_unlock(&ctx->gc_seq_lock);
1197
1198         *vp++ = cpu_to_le32(PTLRPC_SEC_GSS_VERSION);    /* version */
1199         *vp++ = cpu_to_le32(PTLRPCS_FLVR_KRB5P);        /* subflavor */
1200         *vp++ = cpu_to_le32(ctx->gc_proc);              /* proc */
1201         *vp++ = cpu_to_le32(seqnum);                    /* seq */
1202         *vp++ = cpu_to_le32(PTLRPCS_GSS_SVC_PRIVACY);   /* service */
1203         vlen -= 5 * 4;
1204
1205         if (rawobj_serialize(&ctx->gc_wire_ctx, &vp, &vlen)) {
1206                 rc = -EIO;
1207                 goto out;
1208         }
1209         CDEBUG(D_SEC, "encoded wire_ctx length %d\n", ctx->gc_wire_ctx.len);
1210
1211         vpsave = vp++;  /* reserve for size */
1212         vlen -= 4;
1213
1214         msg_buf.buf = (__u8 *) req->rq_reqmsg - GSS_PRIVBUF_PREFIX_LEN;
1215         msg_buf.buflen = req->rq_reqlen + GSS_PRIVBUF_PREFIX_LEN +
1216                                           GSS_PRIVBUF_SUFFIX_LEN;
1217         msg_buf.dataoff = GSS_PRIVBUF_PREFIX_LEN;
1218         msg_buf.datalen = req->rq_reqlen;
1219
1220         cipher_buf.data = (__u8 *) vp;
1221         cipher_buf.len = vlen;
1222
1223         major = kgss_wrap(ctx->gc_gss_ctx, GSS_C_QOP_DEFAULT,
1224                           &msg_buf, &cipher_buf);
1225         if (major) {
1226                 CERROR("cred %p: error wrap: major 0x%x\n", cred, major);
1227                 GOTO(out, rc = -EINVAL);
1228         }
1229
1230         *vpsave = cpu_to_le32(cipher_buf.len);
1231
1232         seclen = seclen - vlen + cipher_buf.len;
1233         sec_hdr = buf_to_sec_hdr(req->rq_reqbuf);
1234         sec_hdr->sec_len = cpu_to_le32(seclen);
1235         req->rq_reqdata_len += size_round(seclen);
1236
1237         CDEBUG(D_SEC, "msg size %d, total sec size %d\n",
1238                req->rq_reqlen, seclen);
1239 out:
1240         gss_put_ctx(ctx);
1241         RETURN(rc);
1242 }
1243
1244 static int gss_cred_unseal(struct ptlrpc_cred *cred,
1245                            struct ptlrpc_request *req)
1246 {
1247         struct gss_cred        *gcred;
1248         struct gss_cl_ctx      *ctx;
1249         struct ptlrpcs_wire_hdr *sec_hdr;
1250         rawobj_t                cipher_text, plain_text;
1251         __u32                   *vp, vlen, subflavor, proc, seq, svc;
1252         __u32                   major, rc;
1253         ENTRY;
1254
1255         LASSERT(req->rq_repbuf);
1256         LASSERT(req->rq_cred == cred);
1257
1258         sec_hdr = buf_to_sec_hdr(req->rq_repbuf);
1259         if (sec_hdr->msg_len != 0) {
1260                 CERROR("unexpected msg_len %u\n", sec_hdr->msg_len);
1261                 RETURN(-EPROTO);
1262         }
1263
1264         vp = (__u32 *) (req->rq_repbuf + sizeof(*sec_hdr));
1265         vlen = sec_hdr->sec_len;
1266
1267         if (vlen < 7 * 4) {
1268                 CERROR("reply sec size %u too small\n", vlen);
1269                 RETURN(-EPROTO);
1270         }
1271
1272         if (*vp++ != cpu_to_le32(PTLRPC_SEC_GSS_VERSION)) {
1273                 CERROR("reply have different gss version\n");
1274                 RETURN(-EPROTO);
1275         }
1276         subflavor = le32_to_cpu(*vp++);
1277         proc = le32_to_cpu(*vp++);
1278         seq = le32_to_cpu(*vp++);
1279         svc = le32_to_cpu(*vp++);
1280         vlen -= 5 * 4;
1281
1282         switch (proc) {
1283         case PTLRPCS_GSS_PROC_DATA:
1284                 if (svc != PTLRPCS_GSS_SVC_PRIVACY) {
1285                         CERROR("Unknown svc %d\n", svc);
1286                         RETURN(-EPROTO);
1287                 }
1288                 if (*vp++ != 0) {
1289                         CERROR("Unexpected ctx handle\n");
1290                         RETURN(-EPROTO);
1291                 }
1292                 vlen -= 4;
1293
1294                 cipher_text.len = le32_to_cpu(*vp++);
1295                 cipher_text.data = (__u8 *) vp;
1296                 vlen -= 4;
1297
1298                 if (vlen < cipher_text.len) {
1299                         CERROR("cipher text to be %u while buf only %u\n",
1300                                 cipher_text.len, vlen);
1301                         RETURN(-EPROTO);
1302                 }
1303
1304                 plain_text = cipher_text;
1305
1306                 gcred = container_of(cred, struct gss_cred, gc_base);
1307                 ctx = gss_cred_get_ctx(cred);
1308                 LASSERT(ctx);
1309
1310                 major = kgss_unwrap(ctx->gc_gss_ctx, GSS_C_QOP_DEFAULT,
1311                                     &cipher_text, &plain_text);
1312                 if (major) {
1313                         CERROR("cred %p: error unwrap: major 0x%x\n",
1314                                cred, major);
1315
1316                         if (major == GSS_S_CREDENTIALS_EXPIRED ||
1317                             major == GSS_S_CONTEXT_EXPIRED) {
1318                                 ptlrpcs_cred_expire(cred);
1319                                 req->rq_ptlrpcs_restart = 1;
1320                                 rc = 0;
1321                         } else
1322                                 rc = -EINVAL;
1323
1324                         GOTO(proc_out, rc);
1325                 }
1326
1327                 req->rq_repmsg = (struct lustre_msg *) vp;
1328                 req->rq_replen = plain_text.len;
1329
1330                 rc = 0;
1331 proc_out:
1332                 gss_put_ctx(ctx);
1333                 break;
1334         default:
1335                 CERROR("unknown gss proc %d\n", proc);
1336                 rc = -EPROTO;
1337         }
1338
1339         RETURN(rc);
1340 }
1341
1342 static void destroy_gss_context(struct ptlrpc_cred *cred)
1343 {
1344         struct ptlrpcs_wire_hdr *hdr;
1345         struct lustre_msg       *lmsg;
1346         struct gss_cred         *gcred;
1347         struct ptlrpc_request    req;
1348         struct obd_import       *imp;
1349         __u32                   *vp, lmsg_size;
1350         struct ptlrpc_request   *raw_req = NULL;
1351         const int                repbuf_len = 256;
1352         char                    *repbuf;
1353         int                      replen, rc;
1354         ENTRY;
1355
1356         imp = cred->pc_sec->ps_import;
1357         LASSERT(imp);
1358
1359         if (test_bit(PTLRPC_CRED_ERROR_BIT, &cred->pc_flags) ||
1360             !test_bit(PTLRPC_CRED_UPTODATE_BIT, &cred->pc_flags)) {
1361                 CDEBUG(D_SEC, "Destroy dead cred %p(%u@%s)\n",
1362                        cred, cred->pc_uid, imp->imp_target_uuid.uuid);
1363                 EXIT;
1364                 return;
1365         }
1366
1367         might_sleep();
1368
1369         /* cred's refcount is 0, steal one */
1370         atomic_inc(&cred->pc_refcount);
1371
1372         gcred = container_of(cred, struct gss_cred, gc_base);
1373         gcred->gc_ctx->gc_proc = PTLRPCS_GSS_PROC_DESTROY;
1374
1375         CDEBUG(D_SEC, "client destroy gss cred %p(%u@%s)\n",
1376                gcred, cred->pc_uid, imp->imp_target_uuid.uuid);
1377
1378         lmsg_size = lustre_msg_size(0, NULL);
1379         req.rq_req_secflvr = cred->pc_sec->ps_flavor;
1380         req.rq_cred = cred;
1381         req.rq_reqbuf_len = sizeof(*hdr) + lmsg_size +
1382                             ptlrpcs_est_req_payload(&req, lmsg_size);
1383
1384         OBD_ALLOC(req.rq_reqbuf, req.rq_reqbuf_len);
1385         if (!req.rq_reqbuf) {
1386                 CERROR("Fail to alloc reqbuf, cancel anyway\n");
1387                 atomic_dec(&cred->pc_refcount);
1388                 EXIT;
1389                 return;
1390         }
1391
1392         /* wire hdr */
1393         hdr = buf_to_sec_hdr(req.rq_reqbuf);
1394         hdr->flavor  = cpu_to_le32(PTLRPCS_FLVR_GSS_AUTH);
1395         hdr->msg_len = cpu_to_le32(lmsg_size);
1396         hdr->sec_len = cpu_to_le32(0);
1397
1398         /* lustre message */
1399         lmsg = buf_to_lustre_msg(req.rq_reqbuf);
1400         lustre_init_msg(lmsg, 0, NULL, NULL);
1401         lmsg->handle   = imp->imp_remote_handle;
1402         lmsg->type     = PTL_RPC_MSG_REQUEST;
1403         lmsg->opc      = SEC_FINI;
1404         lmsg->flags    = 0;
1405         lmsg->conn_cnt = imp->imp_conn_cnt;
1406         /* add this for randomize */
1407         get_random_bytes(&lmsg->last_xid, sizeof(lmsg->last_xid));
1408         get_random_bytes(&lmsg->transno, sizeof(lmsg->transno));
1409
1410         vp = (__u32 *) req.rq_reqbuf;
1411
1412         req.rq_cred = cred;
1413         req.rq_reqmsg = buf_to_lustre_msg(req.rq_reqbuf);
1414         req.rq_reqlen = lmsg_size;
1415         req.rq_reqdata_len = sizeof(*hdr) + lmsg_size;
1416
1417         if (gss_cred_sign(cred, &req)) {
1418                 CERROR("failed to sign, cancel anyway\n");
1419                 atomic_dec(&cred->pc_refcount);
1420                 goto exit;
1421         }
1422         atomic_dec(&cred->pc_refcount);
1423
1424         OBD_ALLOC(repbuf, repbuf_len);
1425         if (!repbuf)
1426                 goto exit;
1427
1428         raw_req = ptl_do_rawrpc(imp, req.rq_reqbuf, req.rq_reqbuf_len,
1429                                 req.rq_reqdata_len, repbuf, repbuf_len, &replen,
1430                                 SECFINI_RPC_TIMEOUT, &rc);
1431         if (!raw_req)
1432                 OBD_FREE(repbuf, repbuf_len);
1433
1434 exit:
1435         if (raw_req == NULL)
1436                 OBD_FREE(req.rq_reqbuf, req.rq_reqbuf_len);
1437         else
1438                 rawrpc_req_finished(raw_req);
1439         EXIT;
1440 }
1441
1442 static void gss_cred_destroy(struct ptlrpc_cred *cred)
1443 {
1444         struct gss_cred *gcred;
1445         ENTRY;
1446
1447         LASSERT(cred);
1448         LASSERT(!atomic_read(&cred->pc_refcount));
1449
1450         gcred = container_of(cred, struct gss_cred, gc_base);
1451         if (gcred->gc_ctx) {
1452                 destroy_gss_context(cred);
1453                 gss_put_ctx(gcred->gc_ctx);
1454         }
1455
1456         CDEBUG(D_SEC, "sec.gss %p: destroy cred %p\n", cred->pc_sec, gcred);
1457
1458         OBD_FREE(gcred, sizeof(*gcred));
1459         EXIT;
1460 }
1461
1462 static struct ptlrpc_credops gss_credops = {
1463         .refresh        = gss_cred_refresh,
1464         .match          = gss_cred_match,
1465         .sign           = gss_cred_sign,
1466         .verify         = gss_cred_verify,
1467         .seal           = gss_cred_seal,
1468         .unseal         = gss_cred_unseal,
1469         .destroy        = gss_cred_destroy,
1470 };
1471
1472 #ifdef __KERNEL__
1473 /*******************************************
1474  * rpc_pipe APIs                           *
1475  *******************************************/
1476 static ssize_t
1477 gss_pipe_upcall(struct file *filp, struct rpc_pipe_msg *msg,
1478                 char *dst, size_t buflen)
1479 {
1480         char *data = (char *)msg->data + msg->copied;
1481         ssize_t mlen = msg->len;
1482         ssize_t left;
1483         ENTRY;
1484
1485         if (mlen > buflen)
1486                 mlen = buflen;
1487         left = copy_to_user(dst, data, mlen);
1488         if (left < 0) {
1489                 msg->errno = left;
1490                 RETURN(left);
1491         }
1492         mlen -= left;
1493         msg->copied += mlen;
1494         msg->errno = 0;
1495         RETURN(mlen);
1496 }
1497
1498 static ssize_t
1499 gss_pipe_downcall(struct file *filp, const char *src, size_t mlen)
1500 {
1501         char *buf;
1502         const int bufsize = 1024;
1503         rawobj_t obj;
1504         struct inode *inode = filp->f_dentry->d_inode;
1505         struct rpc_inode *rpci = RPC_I(inode);
1506         struct obd_import *import;
1507         struct ptlrpc_sec *sec;
1508         struct gss_sec *gsec;
1509         char *obdname;
1510         struct gss_api_mech *mech;
1511         struct vfs_cred vcred = { 0 };
1512         struct ptlrpc_cred *cred;
1513         struct gss_upcall_msg *gss_msg;
1514         struct gss_upcall_msg_data gmd = { 0 };
1515         struct gss_cl_ctx *ctx = NULL;
1516         ssize_t left;
1517         int err, gss_err;
1518         ENTRY;
1519
1520         if (mlen > bufsize) {
1521                 CERROR("mlen %ld > bufsize %d\n", (long)mlen, bufsize);
1522                 RETURN(-ENOSPC);
1523         }
1524
1525         OBD_ALLOC(buf, bufsize);
1526         if (!buf) {
1527                 CERROR("alloc mem failed\n");
1528                 RETURN(-ENOMEM);
1529         }
1530
1531         left = copy_from_user(buf, src, mlen);
1532         if (left)
1533                 GOTO(err_free, err = -EFAULT);
1534
1535         obj.data = (unsigned char *)buf;
1536         obj.len = mlen;
1537
1538         LASSERT(rpci->private);
1539         gsec = (struct gss_sec *)rpci->private;
1540         sec = &gsec->gs_base;
1541         LASSERT(sec->ps_import);
1542         import = class_import_get(sec->ps_import);
1543         LASSERT(import->imp_obd);
1544         obdname = import->imp_obd->obd_name;
1545         mech = gsec->gs_mech;
1546
1547         err = gss_parse_init_downcall(mech, &obj, &ctx, &gmd, &gss_err);
1548         if (err)
1549                 CERROR("parse init downcall err %d\n", err);
1550
1551         vcred.vc_uid = gmd.gum_uid;
1552         vcred.vc_pag = vcred.vc_uid; /* FIXME */
1553
1554         cred = ptlrpcs_cred_lookup(sec, &vcred);
1555         if (!cred) {
1556                 CWARN("didn't find cred for uid %u\n", vcred.vc_uid);
1557                 GOTO(err, err = -EINVAL);
1558         }
1559
1560         if (err || gss_err) {
1561                 set_bit(PTLRPC_CRED_DEAD_BIT, &cred->pc_flags);
1562                 if (err != -ERESTART || gss_err != 0)
1563                         set_bit(PTLRPC_CRED_ERROR_BIT, &cred->pc_flags);
1564                 CERROR("cred %p: rpc err %d, gss err 0x%x, fatal %d\n",
1565                        cred, err, gss_err,
1566                        (test_bit(PTLRPC_CRED_ERROR_BIT, &cred->pc_flags) != 0));
1567         } else {
1568                 CDEBUG(D_SEC, "get initial ctx:\n");
1569                 gss_cred_set_ctx(cred, ctx);
1570         }
1571
1572         spin_lock(&gsec->gs_lock);
1573         gss_msg = gss_find_upcall(gsec, obdname, &gmd);
1574         if (gss_msg) {
1575                 gss_unhash_msg_nolock(gss_msg);
1576                 spin_unlock(&gsec->gs_lock);
1577                 gss_release_msg(gss_msg);
1578         } else
1579                 spin_unlock(&gsec->gs_lock);
1580
1581         ptlrpcs_cred_put(cred, 1);
1582         class_import_put(import);
1583         OBD_FREE(buf, bufsize);
1584         RETURN(mlen);
1585 err:
1586         if (ctx)
1587                 gss_destroy_ctx(ctx);
1588         class_import_put(import);
1589 err_free:
1590         OBD_FREE(buf, bufsize);
1591         CDEBUG(D_SEC, "gss_pipe_downcall returning %d\n", err);
1592         RETURN(err);
1593 }
1594
1595 static
1596 void gss_pipe_destroy_msg(struct rpc_pipe_msg *msg)
1597 {
1598         struct gss_upcall_msg *gmsg;
1599         static unsigned long ratelimit;
1600         ENTRY;
1601
1602         LASSERT(list_empty(&msg->list));
1603
1604         if (msg->errno >= 0) {
1605                 EXIT;
1606                 return;
1607         }
1608
1609         gmsg = container_of(msg, struct gss_upcall_msg, gum_base);
1610         CDEBUG(D_SEC, "destroy gmsg %p\n", gmsg);
1611         LASSERT(atomic_read(&gmsg->gum_refcount) > 0);
1612         atomic_inc(&gmsg->gum_refcount);
1613         gss_unhash_msg(gmsg);
1614         if (msg->errno == -ETIMEDOUT || msg->errno == -EPIPE) {
1615                 unsigned long now = get_seconds();
1616                 if (time_after(now, ratelimit)) {
1617                         CWARN("sec.gss upcall timed out.\n"
1618                               "Please check user daemon is running!\n");
1619                         ratelimit = now + 15;
1620                 }
1621         }
1622         gss_release_msg(gmsg);
1623         EXIT;
1624 }
1625
1626 static
1627 void gss_pipe_release(struct inode *inode)
1628 {
1629         struct rpc_inode *rpci = RPC_I(inode);
1630         struct ptlrpc_sec *sec;
1631         struct gss_sec *gsec;
1632         ENTRY;
1633
1634         gsec = (struct gss_sec *)rpci->private;
1635         sec = &gsec->gs_base;
1636         spin_lock(&gsec->gs_lock);
1637         while (!list_empty(&gsec->gs_upcalls)) {
1638                 struct gss_upcall_msg *gmsg;
1639
1640                 gmsg = list_entry(gsec->gs_upcalls.next,
1641                                   struct gss_upcall_msg, gum_list);
1642                 LASSERT(list_empty(&gmsg->gum_base.list));
1643                 gmsg->gum_base.errno = -EPIPE;
1644                 atomic_inc(&gmsg->gum_refcount);
1645                 gss_unhash_msg_nolock(gmsg);
1646                 gss_release_msg(gmsg);
1647         }
1648         spin_unlock(&gsec->gs_lock);
1649         EXIT;
1650 }
1651
1652 static struct rpc_pipe_ops gss_upcall_ops = {
1653         .upcall         = gss_pipe_upcall,
1654         .downcall       = gss_pipe_downcall,
1655         .destroy_msg    = gss_pipe_destroy_msg,
1656         .release_pipe   = gss_pipe_release,
1657 };
1658 #endif /* __KERNEL__ */
1659
1660 /*********************************************
1661  * GSS security APIs                         *
1662  *********************************************/
1663
1664 static
1665 struct ptlrpc_sec* gss_create_sec(__u32 flavor,
1666                                   const char *pipe_dir,
1667                                   void *pipe_data)
1668 {
1669         struct gss_sec *gsec;
1670         struct ptlrpc_sec *sec;
1671 #ifdef __KERNEL__
1672         char *pos;
1673         int   pipepath_len;
1674         uid_t save_uid;
1675 #endif
1676         ENTRY;
1677
1678         LASSERT(SEC_FLAVOR_MAJOR(flavor) == PTLRPCS_FLVR_MAJOR_GSS);
1679
1680         OBD_ALLOC(gsec, sizeof(*gsec));
1681         if (!gsec) {
1682                 CERROR("can't alloc gsec\n");
1683                 RETURN(NULL);
1684         }
1685
1686         gsec->gs_mech = kgss_subflavor_to_mech(SEC_FLAVOR_SUB(flavor));
1687         if (!gsec->gs_mech) {
1688                 CERROR("subflavor 0x%x not found\n", flavor);
1689                 goto err_free;
1690         }
1691
1692         /* initialize gss sec */
1693 #ifdef __KERNEL__
1694         INIT_LIST_HEAD(&gsec->gs_upcalls);
1695         spin_lock_init(&gsec->gs_lock);
1696
1697         pipepath_len = strlen(LUSTRE_PIPEDIR) + strlen(pipe_dir) +
1698                        strlen(gsec->gs_mech->gm_name) + 3;
1699         OBD_ALLOC(gsec->gs_pipepath, pipepath_len);
1700         if (!gsec->gs_pipepath)
1701                 goto err_mech_put;
1702
1703         /* pipe rpc require root permission */
1704         save_uid = current->fsuid;
1705         current->fsuid = 0;
1706
1707         sprintf(gsec->gs_pipepath, LUSTRE_PIPEDIR"/%s", pipe_dir);
1708         if (IS_ERR(rpc_mkdir(gsec->gs_pipepath, NULL))) {
1709                 CERROR("can't make pipedir %s\n", gsec->gs_pipepath);
1710                 goto err_free_path;
1711         }
1712
1713         sprintf(gsec->gs_pipepath, LUSTRE_PIPEDIR"/%s/%s", pipe_dir,
1714                 gsec->gs_mech->gm_name);
1715         gsec->gs_depipe = rpc_mkpipe(gsec->gs_pipepath, gsec,
1716                                      &gss_upcall_ops, RPC_PIPE_WAIT_FOR_OPEN);
1717         if (IS_ERR(gsec->gs_depipe)) {
1718                 CERROR("failed to make rpc_pipe %s: %ld\n",
1719                         gsec->gs_pipepath, PTR_ERR(gsec->gs_depipe));
1720                 goto err_rmdir;
1721         }
1722         CDEBUG(D_SEC, "gss sec %p, pipe path %s\n", gsec, gsec->gs_pipepath);
1723 #endif
1724
1725         sec = &gsec->gs_base;
1726         sec->ps_expire = GSS_CREDCACHE_EXPIRE;
1727         sec->ps_nextgc = get_seconds() + sec->ps_expire;
1728         sec->ps_flags = 0;
1729
1730         current->fsuid = save_uid;
1731
1732         CDEBUG(D_SEC, "Create sec.gss %p\n", gsec);
1733         RETURN(sec);
1734
1735 #ifdef __KERNEL__
1736 err_rmdir:
1737         pos = strrchr(gsec->gs_pipepath, '/');
1738         LASSERT(pos);
1739         *pos = 0;
1740         rpc_rmdir(gsec->gs_pipepath);
1741 err_free_path:
1742         current->fsuid = save_uid;
1743         OBD_FREE(gsec->gs_pipepath, pipepath_len);
1744 err_mech_put:
1745 #endif
1746         kgss_mech_put(gsec->gs_mech);
1747 err_free:
1748         OBD_FREE(gsec, sizeof(*gsec));
1749         RETURN(NULL);
1750 }
1751
1752 static
1753 void gss_destroy_sec(struct ptlrpc_sec *sec)
1754 {
1755         struct gss_sec *gsec;
1756 #ifdef __KERNEL__
1757         char *pos;
1758         int   pipepath_len;
1759 #endif
1760         ENTRY;
1761
1762         gsec = container_of(sec, struct gss_sec, gs_base);
1763         CDEBUG(D_SEC, "Destroy sec.gss %p\n", gsec);
1764
1765         LASSERT(gsec->gs_mech);
1766         LASSERT(!atomic_read(&sec->ps_refcount));
1767         LASSERT(!atomic_read(&sec->ps_credcount));
1768 #ifdef __KERNEL__
1769         pipepath_len = strlen(gsec->gs_pipepath) + 1;
1770         rpc_unlink(gsec->gs_pipepath);
1771         pos = strrchr(gsec->gs_pipepath, '/');
1772         LASSERT(pos);
1773         *pos = 0;
1774         rpc_rmdir(gsec->gs_pipepath);
1775         OBD_FREE(gsec->gs_pipepath, pipepath_len);
1776 #endif
1777
1778         kgss_mech_put(gsec->gs_mech);
1779         OBD_FREE(gsec, sizeof(*gsec));
1780         EXIT;
1781 }
1782
1783 static
1784 struct ptlrpc_cred * gss_create_cred(struct ptlrpc_sec *sec,
1785                                      struct vfs_cred *vcred)
1786 {
1787         struct gss_cred *gcred;
1788         struct ptlrpc_cred *cred;
1789         ENTRY;
1790
1791         OBD_ALLOC(gcred, sizeof(*gcred));
1792         if (!gcred)
1793                 RETURN(NULL);
1794
1795         cred = &gcred->gc_base;
1796         INIT_LIST_HEAD(&cred->pc_hash);
1797         atomic_set(&cred->pc_refcount, 0);
1798         cred->pc_sec = sec;
1799         cred->pc_ops = &gss_credops;
1800         cred->pc_expire = 0;
1801         cred->pc_flags = 0;
1802         cred->pc_pag = vcred->vc_pag;
1803         cred->pc_uid = vcred->vc_uid;
1804         CDEBUG(D_SEC, "create a gss cred at %p("LPU64"/%u)\n",
1805                cred, vcred->vc_pag, vcred->vc_uid);
1806
1807         RETURN(cred);
1808 }
1809
1810 static int gss_estimate_payload(struct ptlrpc_sec *sec,
1811                                 struct ptlrpc_request *req,
1812                                 int msgsize)
1813 {
1814         switch (SEC_FLAVOR_SVC(req->rq_req_secflvr)) {
1815         case PTLRPCS_SVC_AUTH:
1816                 return GSS_MAX_AUTH_PAYLOAD;
1817         case PTLRPCS_SVC_PRIV:
1818                 return size_round16(GSS_MAX_AUTH_PAYLOAD + msgsize +
1819                                     GSS_PRIVBUF_PREFIX_LEN +
1820                                     GSS_PRIVBUF_SUFFIX_LEN);
1821         default:
1822                 LBUG();
1823                 return 0;
1824         }
1825 }
1826
1827 static int gss_alloc_reqbuf(struct ptlrpc_sec *sec,
1828                             struct ptlrpc_request *req,
1829                             int lmsg_size)
1830 {
1831         int msg_payload, sec_payload;
1832         int privacy, rc;
1833         ENTRY;
1834
1835         /* In PRIVACY mode, lustre message is always 0 (already encoded into
1836          * security payload).
1837          */
1838         privacy = (SEC_FLAVOR_SVC(req->rq_req_secflvr) == PTLRPCS_SVC_PRIV);
1839         msg_payload = privacy ? 0 : lmsg_size;
1840         sec_payload = gss_estimate_payload(sec, req, lmsg_size);
1841
1842         rc = sec_alloc_reqbuf(sec, req, msg_payload, sec_payload);
1843         if (rc)
1844                 return rc;
1845
1846         if (privacy) {
1847                 int buflen = lmsg_size + GSS_PRIVBUF_PREFIX_LEN +
1848                              GSS_PRIVBUF_SUFFIX_LEN;
1849                 char *buf;
1850
1851                 OBD_ALLOC(buf, buflen);
1852                 if (!buf) {
1853                         CERROR("Fail to alloc %d\n", buflen);
1854                         sec_free_reqbuf(sec, req);
1855                         RETURN(-ENOMEM);
1856                 }
1857                 req->rq_reqmsg = (struct lustre_msg *)
1858                                         (buf + GSS_PRIVBUF_PREFIX_LEN);
1859         }
1860
1861         RETURN(0);
1862 }
1863
1864 static void gss_free_reqbuf(struct ptlrpc_sec *sec,
1865                             struct ptlrpc_request *req)
1866 {
1867         char *buf;
1868         int privacy;
1869         ENTRY;
1870
1871         LASSERT(req->rq_reqmsg);
1872         LASSERT(req->rq_reqlen);
1873
1874         privacy = SEC_FLAVOR_SVC(req->rq_req_secflvr) == PTLRPCS_SVC_PRIV;
1875         if (privacy) {
1876                 buf = (char *) req->rq_reqmsg - GSS_PRIVBUF_PREFIX_LEN;
1877                 LASSERT(buf < req->rq_reqbuf ||
1878                         buf >= req->rq_reqbuf + req->rq_reqbuf_len);
1879                 OBD_FREE(buf, req->rq_reqlen + GSS_PRIVBUF_PREFIX_LEN +
1880                               GSS_PRIVBUF_SUFFIX_LEN);
1881                 req->rq_reqmsg = NULL;
1882         }
1883
1884         sec_free_reqbuf(sec, req);
1885 }
1886
1887 static struct ptlrpc_secops gss_secops = {
1888         .create_sec             = gss_create_sec,
1889         .destroy_sec            = gss_destroy_sec,
1890         .create_cred            = gss_create_cred,
1891         .est_req_payload        = gss_estimate_payload,
1892         .est_rep_payload        = gss_estimate_payload,
1893         .alloc_reqbuf           = gss_alloc_reqbuf,
1894         .free_reqbuf            = gss_free_reqbuf,
1895 };
1896
1897 static struct ptlrpc_sec_type gss_type = {
1898         .pst_owner      = THIS_MODULE,
1899         .pst_name       = "sec.gss",
1900         .pst_inst       = ATOMIC_INIT(0),
1901         .pst_flavor     = PTLRPCS_FLVR_MAJOR_GSS,
1902         .pst_ops        = &gss_secops,
1903 };
1904
1905 extern int
1906 (*lustre_secinit_downcall_handler)(char *buffer, unsigned long count);
1907
1908 int __init ptlrpcs_gss_init(void)
1909 {
1910         int rc;
1911
1912         rc = ptlrpcs_register(&gss_type);
1913         if (rc)
1914                 return rc;
1915
1916 #ifdef __KERNEL__
1917         gss_svc_init();
1918
1919         rc = PTR_ERR(rpc_mkdir(LUSTRE_PIPEDIR, NULL));
1920         if (IS_ERR((void *)rc) && rc != -EEXIST) {
1921                 CERROR("fail to make rpcpipedir for lustre\n");
1922                 gss_svc_exit();
1923                 ptlrpcs_unregister(&gss_type);
1924                 return -1;
1925         }
1926         rc = 0;
1927 #else
1928 #endif
1929         rc = init_kerberos_module();
1930         if (rc) {
1931                 ptlrpcs_unregister(&gss_type);
1932         }
1933
1934         lustre_secinit_downcall_handler = gss_send_secinit_rpc;
1935
1936         return rc;
1937 }
1938
1939 #ifdef __KERNEL__
1940 static void __exit ptlrpcs_gss_exit(void)
1941 {
1942         lustre_secinit_downcall_handler = NULL;
1943
1944         cleanup_kerberos_module();
1945         rpc_rmdir(LUSTRE_PIPEDIR);
1946         gss_svc_exit();
1947         ptlrpcs_unregister(&gss_type);
1948 }
1949 #endif
1950
1951 MODULE_AUTHOR("Cluster File Systems, Inc. <info@clusterfs.com>");
1952 MODULE_DESCRIPTION("GSS Security module for Lustre");
1953 MODULE_LICENSE("GPL");
1954
1955 module_init(ptlrpcs_gss_init);
1956 module_exit(ptlrpcs_gss_exit);