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 #if defined(CONFIG_SMP)
 454         LASSERT(spin_is_locked(&gmsg->gum_gsec->gs_lock));
 455 #endif
 456
 457         if (list_empty(&gmsg->gum_list))
 458                 return;
 459
 460         list_del_init(&gmsg->gum_list);
 461         wake_up(&gmsg->gum_waitq);
 462         LASSERT(atomic_read(&gmsg->gum_refcount) > 1);
 463         atomic_dec(&gmsg->gum_refcount);
 464 }
 465
 466 static void
 467 gss_unhash_msg(struct gss_upcall_msg *gmsg)
 468 {
 469         struct gss_sec *gsec = gmsg->gum_gsec;
 470
 471         spin_lock(&gsec->gs_lock);
 472         gss_unhash_msg_nolock(gmsg);
 473         spin_unlock(&gsec->gs_lock);
 474 }
 475
 476 static
 477 struct gss_upcall_msg * gss_find_upcall(struct gss_sec *gsec,
 478                                         char *obdname,
 479                                         struct gss_upcall_msg_data *gmd)
 480 {
 481         struct gss_upcall_msg *gmsg;
 482         ENTRY;
 483
 484 #if defined(CONFIG_SMP)
 485         LASSERT(spin_is_locked(&gsec->gs_lock));
 486 #endif
 487
 488         list_for_each_entry(gmsg, &gsec->gs_upcalls, gum_list) {
 489                 if (memcmp(&gmsg->gum_data, gmd, sizeof(*gmd)))
 490                         continue;
 491                 if (strcmp(gmsg->gum_obdname, obdname))
 492                         continue;
 493                 LASSERT(atomic_read(&gmsg->gum_refcount) > 0);
 494                 atomic_inc(&gmsg->gum_refcount);
 495                 CDEBUG(D_SEC, "found gmsg at %p: obdname %s, uid %d, ref %d\n",
 496                        gmsg, obdname, gmd->gum_uid,
 497                        atomic_read(&gmsg->gum_refcount));
 498                 RETURN(gmsg);
 499         }
 500         RETURN(NULL);
 501 }
 502
 503 static void gss_init_upcall_msg(struct gss_upcall_msg *gmsg,
 504                                 struct gss_sec *gsec, char *obdname,
 505                                 struct gss_upcall_msg_data *gmd)
 506 {
 507         struct rpc_pipe_msg *rpcmsg;
 508         ENTRY;
 509
 510         /* 2 refs: 1 for hash, 1 for current user */
 511         init_waitqueue_head(&gmsg->gum_waitq);
 512         list_add(&gmsg->gum_list, &gsec->gs_upcalls);
 513         atomic_set(&gmsg->gum_refcount, 2);
 514         gmsg->gum_gsec = gsec;
 515         strncpy(gmsg->gum_obdname, obdname, sizeof(gmsg->gum_obdname));
 516         memcpy(&gmsg->gum_data, gmd, sizeof(*gmd));
 517
 518         rpcmsg = &gmsg->gum_base;
 519         INIT_LIST_HEAD(&rpcmsg->list);
 520         rpcmsg->data = &gmsg->gum_data;
 521         rpcmsg->len = sizeof(gmsg->gum_data);
 522         rpcmsg->copied = 0;
 523         rpcmsg->errno = 0;
 524         EXIT;
 525 }
 526 #endif /* __KERNEL__ */
 527
 528 /********************************************
 529  * gss cred manipulation helpers            *
 530  ********************************************/
 531 #if 0
 532 static
 533 int gss_cred_is_uptodate_ctx(struct ptlrpc_cred *cred)
 534 {
 535         struct gss_cred *gcred = container_of(cred, struct gss_cred, gc_base);
 536         int res = 0;
 537
 538         read_lock(&gss_ctx_lock);
 539         if (((cred->pc_flags & PTLRPC_CRED_FLAGS_MASK) ==
 540              PTLRPC_CRED_UPTODATE) &&
 541             gcred->gc_ctx)
 542                 res = 1;
 543         read_unlock(&gss_ctx_lock);
 544         return res;
 545 }
 546 #endif
 547
 548 static inline
 549 struct gss_cl_ctx * gss_get_ctx(struct gss_cl_ctx *ctx)
 550 {
 551         atomic_inc(&ctx->gc_refcount);
 552         return ctx;
 553 }
 554
 555 static
 556 void gss_destroy_ctx(struct gss_cl_ctx *ctx)
 557 {
 558         ENTRY;
 559
 560         CDEBUG(D_SEC, "destroy cl_ctx %p\n", ctx);
 561         if (ctx->gc_gss_ctx)
 562                 kgss_delete_sec_context(&ctx->gc_gss_ctx);
 563
 564         if (ctx->gc_wire_ctx.len > 0) {
 565                 OBD_FREE(ctx->gc_wire_ctx.data, ctx->gc_wire_ctx.len);
 566                 ctx->gc_wire_ctx.len = 0;
 567         }
 568
 569         OBD_FREE(ctx, sizeof(*ctx));
 570 }
 571
 572 static
 573 void gss_put_ctx(struct gss_cl_ctx *ctx)
 574 {
 575         if (atomic_dec_and_test(&ctx->gc_refcount))
 576                 gss_destroy_ctx(ctx);
 577 }
 578
 579 static
 580 struct gss_cl_ctx *gss_cred_get_ctx(struct ptlrpc_cred *cred)
 581 {
 582         struct gss_cred *gcred = container_of(cred, struct gss_cred, gc_base);
 583         struct gss_cl_ctx *ctx = NULL;
 584
 585         read_lock(&gss_ctx_lock);
 586         if (gcred->gc_ctx)
 587                 ctx = gss_get_ctx(gcred->gc_ctx);
 588         read_unlock(&gss_ctx_lock);
 589         return ctx;
 590 }
 591
 592 static
 593 void gss_cred_set_ctx(struct ptlrpc_cred *cred, struct gss_cl_ctx *ctx)
 594 {
 595         struct gss_cred *gcred = container_of(cred, struct gss_cred, gc_base);
 596         struct gss_cl_ctx *old;
 597         __u64 ctx_expiry;
 598         ENTRY;
 599
 600         if (kgss_inquire_context(ctx->gc_gss_ctx, &ctx_expiry)) {
 601                 CERROR("unable to get expire time\n");
 602                 ctx_expiry = 1; /* make it expired now */
 603         }
 604         cred->pc_expire = (unsigned long) ctx_expiry;
 605
 606         write_lock(&gss_ctx_lock);
 607         old = gcred->gc_ctx;
 608         gcred->gc_ctx = ctx;
 609         set_bit(PTLRPC_CRED_UPTODATE_BIT, &cred->pc_flags);
 610         write_unlock(&gss_ctx_lock);
 611         if (old)
 612                 gss_put_ctx(old);
 613
 614         CDEBUG(D_SEC, "client refreshed gss cred %p(uid %u)\n",
 615                cred, cred->pc_uid);
 616         EXIT;
 617 }
 618
 619 static int
 620 simple_get_bytes(char **buf, __u32 *buflen, void *res, __u32 reslen)
 621 {
 622         if (*buflen < reslen) {
 623                 CERROR("buflen %u < %u\n", *buflen, reslen);
 624                 return -EINVAL;
 625         }
 626
 627         memcpy(res, *buf, reslen);
 628         *buf += reslen;
 629         *buflen -= reslen;
 630         return 0;
 631 }
 632
 633 /* data passed down:
 634  *  - uid
 635  *  - timeout
 636  *  - gc_win / error
 637  *  - wire_ctx (rawobj)
 638  *  - mech_ctx? (rawobj)
 639  */
 640 static
 641 int gss_parse_init_downcall(struct gss_api_mech *gm, rawobj_t *buf,
 642                             struct gss_cl_ctx **gc,
 643                             struct gss_upcall_msg_data *gmd, int *gss_err)
 644 {
 645         char *p = (char *)buf->data;
 646         struct gss_cl_ctx *ctx;
 647         __u32 len = buf->len;
 648         unsigned int timeout;
 649         rawobj_t tmp_buf;
 650         int err = -EPERM;
 651         ENTRY;
 652
 653         *gc = NULL;
 654         *gss_err = 0;
 655
 656         OBD_ALLOC(ctx, sizeof(*ctx));
 657         if (!ctx)
 658                 RETURN(-ENOMEM);
 659
 660         ctx->gc_proc = RPC_GSS_PROC_DATA;
 661         ctx->gc_seq = 0;
 662         spin_lock_init(&ctx->gc_seq_lock);
 663         atomic_set(&ctx->gc_refcount,1);
 664
 665         if (simple_get_bytes(&p, &len, &gmd->gum_uid, sizeof(gmd->gum_uid)))
 666                 goto err_free_ctx;
 667         if (simple_get_bytes(&p, &len, &gmd->gum_svc, sizeof(gmd->gum_svc)))
 668                 goto err_free_ctx;
 669         if (simple_get_bytes(&p, &len, &gmd->gum_nal, sizeof(gmd->gum_nal)))
 670                 goto err_free_ctx;
 671         if (simple_get_bytes(&p, &len, &gmd->gum_netid, sizeof(gmd->gum_netid)))
 672                 goto err_free_ctx;
 673         if (simple_get_bytes(&p, &len, &gmd->gum_nid, sizeof(gmd->gum_nid)))
 674                 goto err_free_ctx;
 675         /* FIXME: discarded timeout for now */
 676         if (simple_get_bytes(&p, &len, &timeout, sizeof(timeout)))
 677                 goto err_free_ctx;
 678         if (simple_get_bytes(&p, &len, &ctx->gc_win, sizeof(ctx->gc_win)))
 679                 goto err_free_ctx;
 680
 681         /* lgssd signals an error by passing ctx->gc_win = 0: */
 682         if (!ctx->gc_win) {
 683                 /* in which case the next 2 int are:
 684                  * - rpc error
 685                  * - gss error
 686                  */
 687                 if (simple_get_bytes(&p, &len, &err, sizeof(err))) {
 688                         err = -EPERM;
 689                         goto err_free_ctx;
 690                 }
 691                 if (simple_get_bytes(&p, &len, gss_err, sizeof(*gss_err))) {
 692                         err = -EPERM;
 693                         goto err_free_ctx;
 694                 }
 695                 if (err == 0 && *gss_err == 0) {
 696                         CERROR("no error passed from downcall\n");
 697                         err = -EPERM;
 698                 }
 699                 goto err_free_ctx;
 700         }
 701
 702         if (rawobj_extract_local(&tmp_buf, (__u32 **) ((void *)&p), &len))
 703                 goto err_free_ctx;
 704         if (rawobj_dup(&ctx->gc_wire_ctx, &tmp_buf)) {
 705                 err = -ENOMEM;
 706                 goto err_free_ctx;
 707         }
 708         if (rawobj_extract_local(&tmp_buf, (__u32 **) ((void *)&p), &len))
 709                 goto err_free_wire_ctx;
 710         if (len) {
 711                 CERROR("unexpected trailing %u bytes\n", len);
 712                 goto err_free_wire_ctx;
 713         }
 714         if (kgss_import_sec_context(&tmp_buf, gm, &ctx->gc_gss_ctx))
 715                 goto err_free_wire_ctx;
 716
 717         *gc = ctx;
 718         RETURN(0);
 719
 720 err_free_wire_ctx:
 721         if (ctx->gc_wire_ctx.data)
 722                 OBD_FREE(ctx->gc_wire_ctx.data, ctx->gc_wire_ctx.len);
 723 err_free_ctx:
 724         OBD_FREE(ctx, sizeof(*ctx));
 725         CDEBUG(D_SEC, "err_code %d, gss code %d\n", err, *gss_err);
 726         return err;
 727 }
 728
 729 /***************************************
 730  * cred APIs                           *
 731  ***************************************/
 732 #ifdef __KERNEL__
 733 static int gss_cred_refresh(struct ptlrpc_cred *cred)
 734 {
 735         struct obd_import          *import;
 736         struct gss_sec             *gsec;
 737         struct gss_upcall_msg      *gss_msg, *gss_new;
 738         struct gss_upcall_msg_data  gmd;
 739         struct dentry              *dentry;
 740         char                       *obdname, *obdtype;
 741         wait_queue_t                wait;
 742         uid_t                       uid = cred->pc_uid;
 743         int                         res;
 744         ENTRY;
 745
 746         might_sleep();
 747
 748         /* any flags means it has been handled, do nothing */
 749         if (cred->pc_flags & PTLRPC_CRED_FLAGS_MASK)
 750                 RETURN(0);
 751
 752         LASSERT(cred->pc_sec);
 753         LASSERT(cred->pc_sec->ps_import);
 754         LASSERT(cred->pc_sec->ps_import->imp_obd);
 755
 756         import = cred->pc_sec->ps_import;
 757         if (!import->imp_connection) {
 758                 CERROR("import has no connection set\n");
 759                 RETURN(-EINVAL);
 760         }
 761
 762         gmd.gum_uid = uid;
 763         gmd.gum_nal = import->imp_connection->c_peer.peer_ni->pni_number;
 764         gmd.gum_netid = 0;
 765         gmd.gum_nid = import->imp_connection->c_peer.peer_id.nid;
 766
 767         obdtype = import->imp_obd->obd_type->typ_name;
 768         if (!strcmp(obdtype, "mdc"))
 769                 gmd.gum_svc = LUSTRE_GSS_SVC_MDS;
 770         else if (!strcmp(obdtype, "osc"))
 771                 gmd.gum_svc = LUSTRE_GSS_SVC_OSS;
 772         else {
 773                 CERROR("gss on %s?\n", obdtype);
 774                 RETURN(-EINVAL);
 775         }
 776
 777         gsec = container_of(cred->pc_sec, struct gss_sec, gs_base);
 778         obdname = import->imp_obd->obd_name;
 779         dentry = gsec->gs_depipe;
 780         gss_new = NULL;
 781         res = 0;
 782
 783         CDEBUG(D_SEC, "Initiate gss context %p(%u@%s)\n",
 784                container_of(cred, struct gss_cred, gc_base),
 785                uid, import->imp_target_uuid.uuid);
 786
 787 again:
 788         spin_lock(&gsec->gs_lock);
 789         gss_msg = gss_find_upcall(gsec, obdname, &gmd);
 790         if (gss_msg) {
 791                 if (gss_new) {
 792                         OBD_FREE(gss_new, sizeof(*gss_new));
 793                         gss_new = NULL;
 794                 }
 795                 GOTO(waiting, res);
 796         }
 797
 798         if (!gss_new) {
 799                 spin_unlock(&gsec->gs_lock);
 800                 OBD_ALLOC(gss_new, sizeof(*gss_new));
 801                 if (!gss_new)
 802                         RETURN(-ENOMEM);
 803                 goto again;
 804         }
 805         /* so far we'v created gss_new */
 806         gss_init_upcall_msg(gss_new, gsec, obdname, &gmd);
 807
 808         /* we'v created upcall msg, nobody else should touch the
 809          * flag of this cred, unless be set as dead/expire by
 810          * administrator via lctl etc.
 811          */
 812         if (cred->pc_flags & PTLRPC_CRED_FLAGS_MASK) {
 813                 CWARN("cred %p("LPU64"/%u) was set flags %lx unexpectedly\n",
 814                       cred, cred->pc_pag, cred->pc_uid, cred->pc_flags);
 815                 cred->pc_flags |= PTLRPC_CRED_DEAD | PTLRPC_CRED_ERROR;
 816                 gss_unhash_msg_nolock(gss_new);
 817                 spin_unlock(&gsec->gs_lock);
 818                 gss_release_msg(gss_new);
 819                 RETURN(0);
 820         }
 821
 822         /* need to make upcall now */
 823         spin_unlock(&gsec->gs_lock);
 824         res = rpc_queue_upcall(dentry->d_inode, &gss_new->gum_base);
 825         if (res) {
 826                 CERROR("rpc_queue_upcall failed: %d\n", res);
 827                 gss_unhash_msg(gss_new);
 828                 gss_release_msg(gss_new);
 829                 cred->pc_flags |= PTLRPC_CRED_DEAD | PTLRPC_CRED_ERROR;
 830                 RETURN(res);
 831         }
 832         gss_msg = gss_new;
 833         spin_lock(&gsec->gs_lock);
 834
 835 waiting:
 836         /* upcall might finish quickly */
 837         if (list_empty(&gss_msg->gum_list)) {
 838                 spin_unlock(&gsec->gs_lock);
 839                 res = 0;
 840                 goto out;
 841         }
 842
 843         init_waitqueue_entry(&wait, current);
 844         set_current_state(TASK_INTERRUPTIBLE);
 845         add_wait_queue(&gss_msg->gum_waitq, &wait);
 846         spin_unlock(&gsec->gs_lock);
 847
 848         if (gss_new)
 849                 res = schedule_timeout(CRED_REFRESH_UPCALL_TIMEOUT * HZ);
 850         else {
 851                 schedule();
 852                 res = 0;
 853         }
 854
 855         remove_wait_queue(&gss_msg->gum_waitq, &wait);
 856
 857         /* - the one who refresh the cred for us should also be responsible
 858          *   to set the status of cred, we can simply return.
 859          * - if cred flags has been set, we also don't need to do that again,
 860          *   no matter signal pending or timeout etc.
 861          */
 862         if (!gss_new || cred->pc_flags & PTLRPC_CRED_FLAGS_MASK)
 863                 goto out;
 864
 865         if (signal_pending(current)) {
 866                 CERROR("%s: cred %p: interrupted upcall\n",
 867                        current->comm, cred);
 868                 cred->pc_flags |= PTLRPC_CRED_DEAD | PTLRPC_CRED_ERROR;
 869                 res = -EINTR;
 870         } else if (res == 0) {
 871                 CERROR("cred %p: upcall timedout\n", cred);
 872                 set_bit(PTLRPC_CRED_DEAD_BIT, &cred->pc_flags);
 873                 res = -ETIMEDOUT;
 874         } else
 875                 res = 0;
 876
 877 out:
 878         gss_release_msg(gss_msg);
 879
 880         RETURN(res);
 881 }
 882 #else /* !__KERNEL__ */
 883 extern int lgss_handle_krb5_upcall(uid_t uid, __u32 dest_ip,
 884                                    char *obd_name,
 885                                    char *buf, int bufsize,
 886                                    int (*callback)(char*, unsigned long));
 887
 888 static int gss_cred_refresh(struct ptlrpc_cred *cred)
 889 {
 890         char                    buf[4096];
 891         rawobj_t                obj;
 892         struct obd_import      *imp;
 893         struct gss_sec         *gsec;
 894         struct gss_api_mech    *mech;
 895         struct gss_cl_ctx      *ctx = NULL;
 896         struct vfs_cred         vcred = { 0 };
 897         ptl_nid_t               peer_nid;
 898         __u32                   dest_ip;
 899         __u32                   subflavor;
 900         int                     rc, gss_err;
 901
 902         LASSERT(cred);
 903         LASSERT(cred->pc_sec);
 904         LASSERT(cred->pc_sec->ps_import);
 905         LASSERT(cred->pc_sec->ps_import->imp_obd);
 906
 907         if (ptlrpcs_cred_is_uptodate(cred))
 908                 RETURN(0);
 909
 910         imp = cred->pc_sec->ps_import;
 911         peer_nid = imp->imp_connection->c_peer.peer_id.nid;
 912         dest_ip = (__u32) (peer_nid & 0xFFFFFFFF);
 913         subflavor = cred->pc_sec->ps_flavor.subflavor;
 914
 915         if (subflavor != PTLRPC_SEC_GSS_KRB5I) {
 916                 CERROR("unknown subflavor %u\n", subflavor);
 917                 GOTO(err_out, rc = -EINVAL);
 918         }
 919
 920         rc = lgss_handle_krb5_upcall(cred->pc_uid, dest_ip,
 921                                      imp->imp_obd->obd_name,
 922                                      buf, sizeof(buf),
 923                                      gss_send_secinit_rpc);
 924         LASSERT(rc != 0);
 925         if (rc < 0)
 926                 goto err_out;
 927
 928         obj.data = buf;
 929         obj.len = rc;
 930
 931         gsec = container_of(cred->pc_sec, struct gss_sec, gs_base);
 932         mech = gsec->gs_mech;
 933         LASSERT(mech);
 934         rc = gss_parse_init_downcall(mech, &obj, &ctx, &vcred, &dest_ip,
 935                                      &gss_err);
 936         if (rc || gss_err) {
 937                 CERROR("parse init downcall: rpc %d, gss 0x%x\n", rc, gss_err);
 938                 if (rc != -ERESTART || gss_err != 0)
 939                         set_bit(PTLRPC_CRED_ERROR_BIT, &cred->pc_flags);
 940                 if (rc == 0)
 941                         rc = -EPERM;
 942                 goto err_out;
 943         }
 944
 945         LASSERT(ctx);
 946         gss_cred_set_ctx(cred, ctx);
 947         LASSERT(gss_cred_is_uptodate_ctx(cred));
 948
 949         return 0;
 950 err_out:
 951         set_bit(PTLRPC_CRED_DEAD_BIT, &cred->pc_flags);
 952         return rc;
 953 }
 954 #endif
 955
 956 static int gss_cred_match(struct ptlrpc_cred *cred,
 957                           struct vfs_cred *vcred)
 958 {
 959         RETURN(cred->pc_pag == vcred->vc_pag);
 960 }
 961
 962 static int gss_cred_sign(struct ptlrpc_cred *cred,
 963                          struct ptlrpc_request *req)
 964 {
 965         struct gss_cred         *gcred;
 966         struct gss_cl_ctx       *ctx;
 967         rawobj_t                 lmsg, mic;
 968         __u32                   *vp, *vpsave, vlen, seclen;
 969         __u32                    seqnum, major, rc = 0;
 970         ENTRY;
 971
 972         LASSERT(req->rq_reqbuf);
 973         LASSERT(req->rq_cred == cred);
 974
 975         gcred = container_of(cred, struct gss_cred, gc_base);
 976         ctx = gss_cred_get_ctx(cred);
 977         if (!ctx) {
 978                 CERROR("cred %p("LPU64"/%u) invalidated?\n",
 979                         cred, cred->pc_pag, cred->pc_uid);
 980                 RETURN(-EPERM);
 981         }
 982
 983         lmsg.len = req->rq_reqlen;
 984         lmsg.data = (__u8 *) req->rq_reqmsg;
 985
 986         vp = (__u32 *) (lmsg.data + lmsg.len);
 987         vlen = req->rq_reqbuf_len - sizeof(struct ptlrpcs_wire_hdr) -
 988                lmsg.len;
 989         seclen = vlen;
 990
 991         if (vlen < 6 * 4 + size_round4(ctx->gc_wire_ctx.len)) {
 992                 CERROR("vlen %d, need %d\n",
 993                         vlen, 6 * 4 + size_round4(ctx->gc_wire_ctx.len));
 994                 rc = -EIO;
 995                 goto out;
 996         }
 997
 998         spin_lock(&ctx->gc_seq_lock);
 999         seqnum = ctx->gc_seq++;
1000         spin_unlock(&ctx->gc_seq_lock);
1001
1002         *vp++ = cpu_to_le32(PTLRPC_SEC_GSS_VERSION);    /* version */
1003         *vp++ = cpu_to_le32(PTLRPCS_FLVR_KRB5I);        /* subflavor */
1004         *vp++ = cpu_to_le32(ctx->gc_proc);              /* proc */
1005         *vp++ = cpu_to_le32(seqnum);                    /* seq */
1006         *vp++ = cpu_to_le32(PTLRPCS_GSS_SVC_INTEGRITY); /* service */
1007         vlen -= 5 * 4;
1008
1009         if (rawobj_serialize(&ctx->gc_wire_ctx, &vp, &vlen)) {
1010                 rc = -EIO;
1011                 goto out;
1012         }
1013         CDEBUG(D_SEC, "encoded wire_ctx length %d\n", ctx->gc_wire_ctx.len);
1014
1015         vpsave = vp++;  /* reserve for size */
1016         vlen -= 4;
1017
1018         mic.len = vlen;
1019         mic.data = (unsigned char *)vp;
1020
1021         CDEBUG(D_SEC, "reqbuf at %p, lmsg at %p, len %d, mic at %p, len %d\n",
1022                req->rq_reqbuf, lmsg.data, lmsg.len, mic.data, mic.len);
1023         major = kgss_get_mic(ctx->gc_gss_ctx, GSS_C_QOP_DEFAULT, &lmsg, &mic);
1024         if (major) {
1025                 CERROR("cred %p: req %p compute mic error, major %x\n",
1026                        cred, req, major);
1027                 rc = -EACCES;
1028                 goto out;
1029         }
1030
1031         *vpsave = cpu_to_le32(mic.len);
1032
1033         seclen = seclen - vlen + mic.len;
1034         buf_to_sec_hdr(req->rq_reqbuf)->sec_len = cpu_to_le32(seclen);
1035         req->rq_reqdata_len += size_round(seclen);
1036         CDEBUG(D_SEC, "msg size %d, checksum size %d, total sec size %d\n",
1037                lmsg.len, mic.len, seclen);
1038 out:
1039         gss_put_ctx(ctx);
1040         RETURN(rc);
1041 }
1042
1043 static int gss_cred_verify(struct ptlrpc_cred *cred,
1044                            struct ptlrpc_request *req)
1045 {
1046         struct gss_cred        *gcred;
1047         struct gss_cl_ctx      *ctx;
1048         struct ptlrpcs_wire_hdr *sec_hdr;
1049         rawobj_t                lmsg, mic;
1050         __u32                   *vp, vlen, subflavor, proc, seq, svc;
1051         __u32                   major, minor, rc;
1052         ENTRY;
1053
1054         LASSERT(req->rq_repbuf);
1055         LASSERT(req->rq_cred == cred);
1056
1057         sec_hdr = buf_to_sec_hdr(req->rq_repbuf);
1058         vp = (__u32 *) (req->rq_repbuf + sizeof(*sec_hdr) + sec_hdr->msg_len);
1059         vlen = sec_hdr->sec_len;
1060
1061         if (vlen < 7 * 4) {
1062                 CERROR("reply sec size %u too small\n", vlen);
1063                 RETURN(-EPROTO);
1064         }
1065
1066         if (*vp++ != cpu_to_le32(PTLRPC_SEC_GSS_VERSION)) {
1067                 CERROR("reply have different gss version\n");
1068                 RETURN(-EPROTO);
1069         }
1070         subflavor = le32_to_cpu(*vp++);
1071         proc = le32_to_cpu(*vp++);
1072         vlen -= 3 * 4;
1073
1074         switch (proc) {
1075         case PTLRPCS_GSS_PROC_DATA:
1076                 seq = le32_to_cpu(*vp++);
1077                 svc = le32_to_cpu(*vp++);
1078                 if (svc != PTLRPCS_GSS_SVC_INTEGRITY) {
1079                         CERROR("Unknown svc %d\n", svc);
1080                         RETURN(-EPROTO);
1081                 }
1082                 if (*vp++ != 0) {
1083                         CERROR("Unexpected ctx handle\n");
1084                         RETURN(-EPROTO);
1085                 }
1086                 mic.len = le32_to_cpu(*vp++);
1087                 vlen -= 4 * 4;
1088                 if (vlen < mic.len) {
1089                         CERROR("vlen %d, mic.len %d\n", vlen, mic.len);
1090                         RETURN(-EINVAL);
1091                 }
1092                 mic.data = (unsigned char *)vp;
1093
1094                 gcred = container_of(cred, struct gss_cred, gc_base);
1095                 ctx = gss_cred_get_ctx(cred);
1096                 LASSERT(ctx);
1097
1098                 lmsg.len = sec_hdr->msg_len;
1099                 lmsg.data = (__u8 *) buf_to_lustre_msg(req->rq_repbuf);
1100
1101                 major = kgss_verify_mic(ctx->gc_gss_ctx, &lmsg, &mic, NULL);
1102                 if (major != GSS_S_COMPLETE) {
1103                         CERROR("cred %p: req %p verify mic error: major %x\n",
1104                                cred, req, major);
1105
1106                         if (major == GSS_S_CREDENTIALS_EXPIRED ||
1107                             major == GSS_S_CONTEXT_EXPIRED) {
1108                                 ptlrpcs_cred_expire(cred);
1109                                 req->rq_ptlrpcs_restart = 1;
1110                                 rc = 0;
1111                         } else
1112                                 rc = -EINVAL;
1113
1114                         GOTO(proc_data_out, rc);
1115                 }
1116
1117                 req->rq_repmsg = (struct lustre_msg *) lmsg.data;
1118                 req->rq_replen = lmsg.len;
1119
1120                 /* here we could check the seq number is the same one
1121                  * we sent to server. but portals has prevent us from
1122                  * replay attack, so maybe we don't need check it again.
1123                  */
1124                 rc = 0;
1125 proc_data_out:
1126                 gss_put_ctx(ctx);
1127                 break;
1128         case PTLRPCS_GSS_PROC_ERR:
1129                 major = le32_to_cpu(*vp++);
1130                 minor = le32_to_cpu(*vp++);
1131                 /* server return NO_CONTEXT might be caused by context expire
1132                  * or server reboot/failover. we refresh the cred transparently
1133                  * to upper layer.
1134                  * In some cases, our gss handle is possible to be incidentally
1135                  * identical to another handle since the handle itself is not
1136                  * fully random. In krb5 case, the GSS_S_BAD_SIG will be
1137                  * returned, maybe other gss error for other mechanism. Here we
1138                  * only consider krb5 mech (FIXME) and try to establish new
1139                  * context.
1140                  */
1141                 if (major == GSS_S_NO_CONTEXT ||
1142                     major == GSS_S_BAD_SIG) {
1143                         CWARN("req %p: server report cred %p %s, expired?\n",
1144                                req, cred, (major == GSS_S_NO_CONTEXT) ?
1145                                            "NO_CONTEXT" : "BAD_SIG");
1146
1147                         ptlrpcs_cred_expire(cred);
1148                         req->rq_ptlrpcs_restart = 1;
1149                         rc = 0;
1150                 } else {
1151                         CERROR("req %p: unrecognized gss error (%x/%x)\n",
1152                                 req, major, minor);
1153                         rc = -EACCES;
1154                 }
1155                 break;
1156         default:
1157                 CERROR("unknown gss proc %d\n", proc);
1158                 rc = -EPROTO;
1159         }
1160
1161         RETURN(rc);
1162 }
1163
1164 static int gss_cred_seal(struct ptlrpc_cred *cred,
1165                          struct ptlrpc_request *req)
1166 {
1167         struct gss_cred         *gcred;
1168         struct gss_cl_ctx       *ctx;
1169         struct ptlrpcs_wire_hdr *sec_hdr;
1170         rawobj_buf_t             msg_buf;
1171         rawobj_t                 cipher_buf;
1172         __u32                   *vp, *vpsave, vlen, seclen;
1173         __u32                    major, seqnum, rc = 0;
1174         ENTRY;
1175
1176         LASSERT(req->rq_reqbuf);
1177         LASSERT(req->rq_cred == cred);
1178
1179         gcred = container_of(cred, struct gss_cred, gc_base);
1180         ctx = gss_cred_get_ctx(cred);
1181         if (!ctx) {
1182                 CERROR("cred %p("LPU64"/%u) invalidated?\n",
1183                         cred, cred->pc_pag, cred->pc_uid);
1184                 RETURN(-EPERM);
1185         }
1186
1187         vp = (__u32 *) (req->rq_reqbuf + sizeof(*sec_hdr));
1188         vlen = req->rq_reqbuf_len - sizeof(*sec_hdr);
1189         seclen = vlen;
1190
1191         if (vlen < 6 * 4 + size_round4(ctx->gc_wire_ctx.len)) {
1192                 CERROR("vlen %d, need %d\n",
1193                         vlen, 6 * 4 + size_round4(ctx->gc_wire_ctx.len));
1194                 rc = -EIO;
1195                 goto out;
1196         }
1197
1198         spin_lock(&ctx->gc_seq_lock);
1199         seqnum = ctx->gc_seq++;
1200         spin_unlock(&ctx->gc_seq_lock);
1201
1202         *vp++ = cpu_to_le32(PTLRPC_SEC_GSS_VERSION);    /* version */
1203         *vp++ = cpu_to_le32(PTLRPCS_FLVR_KRB5P);        /* subflavor */
1204         *vp++ = cpu_to_le32(ctx->gc_proc);              /* proc */
1205         *vp++ = cpu_to_le32(seqnum);                    /* seq */
1206         *vp++ = cpu_to_le32(PTLRPCS_GSS_SVC_PRIVACY);   /* service */
1207         vlen -= 5 * 4;
1208
1209         if (rawobj_serialize(&ctx->gc_wire_ctx, &vp, &vlen)) {
1210                 rc = -EIO;
1211                 goto out;
1212         }
1213         CDEBUG(D_SEC, "encoded wire_ctx length %d\n", ctx->gc_wire_ctx.len);
1214
1215         vpsave = vp++;  /* reserve for size */
1216         vlen -= 4;
1217
1218         msg_buf.buf = (__u8 *) req->rq_reqmsg - GSS_PRIVBUF_PREFIX_LEN;
1219         msg_buf.buflen = req->rq_reqlen + GSS_PRIVBUF_PREFIX_LEN +
1220                                           GSS_PRIVBUF_SUFFIX_LEN;
1221         msg_buf.dataoff = GSS_PRIVBUF_PREFIX_LEN;
1222         msg_buf.datalen = req->rq_reqlen;
1223
1224         cipher_buf.data = (__u8 *) vp;
1225         cipher_buf.len = vlen;
1226
1227         major = kgss_wrap(ctx->gc_gss_ctx, GSS_C_QOP_DEFAULT,
1228                           &msg_buf, &cipher_buf);
1229         if (major) {
1230                 CERROR("cred %p: error wrap: major 0x%x\n", cred, major);
1231                 GOTO(out, rc = -EINVAL);
1232         }
1233
1234         *vpsave = cpu_to_le32(cipher_buf.len);
1235
1236         seclen = seclen - vlen + cipher_buf.len;
1237         sec_hdr = buf_to_sec_hdr(req->rq_reqbuf);
1238         sec_hdr->sec_len = cpu_to_le32(seclen);
1239         req->rq_reqdata_len += size_round(seclen);
1240
1241         CDEBUG(D_SEC, "msg size %d, total sec size %d\n",
1242                req->rq_reqlen, seclen);
1243 out:
1244         gss_put_ctx(ctx);
1245         RETURN(rc);
1246 }
1247
1248 static int gss_cred_unseal(struct ptlrpc_cred *cred,
1249                            struct ptlrpc_request *req)
1250 {
1251         struct gss_cred        *gcred;
1252         struct gss_cl_ctx      *ctx;
1253         struct ptlrpcs_wire_hdr *sec_hdr;
1254         rawobj_t                cipher_text, plain_text;
1255         __u32                   *vp, vlen, subflavor, proc, seq, svc;
1256         __u32                   major, rc;
1257         ENTRY;
1258
1259         LASSERT(req->rq_repbuf);
1260         LASSERT(req->rq_cred == cred);
1261
1262         sec_hdr = buf_to_sec_hdr(req->rq_repbuf);
1263         if (sec_hdr->msg_len != 0) {
1264                 CERROR("unexpected msg_len %u\n", sec_hdr->msg_len);
1265                 RETURN(-EPROTO);
1266         }
1267
1268         vp = (__u32 *) (req->rq_repbuf + sizeof(*sec_hdr));
1269         vlen = sec_hdr->sec_len;
1270
1271         if (vlen < 7 * 4) {
1272                 CERROR("reply sec size %u too small\n", vlen);
1273                 RETURN(-EPROTO);
1274         }
1275
1276         if (*vp++ != cpu_to_le32(PTLRPC_SEC_GSS_VERSION)) {
1277                 CERROR("reply have different gss version\n");
1278                 RETURN(-EPROTO);
1279         }
1280         subflavor = le32_to_cpu(*vp++);
1281         proc = le32_to_cpu(*vp++);
1282         seq = le32_to_cpu(*vp++);
1283         svc = le32_to_cpu(*vp++);
1284         vlen -= 5 * 4;
1285
1286         switch (proc) {
1287         case PTLRPCS_GSS_PROC_DATA:
1288                 if (svc != PTLRPCS_GSS_SVC_PRIVACY) {
1289                         CERROR("Unknown svc %d\n", svc);
1290                         RETURN(-EPROTO);
1291                 }
1292                 if (*vp++ != 0) {
1293                         CERROR("Unexpected ctx handle\n");
1294                         RETURN(-EPROTO);
1295                 }
1296                 vlen -= 4;
1297
1298                 cipher_text.len = le32_to_cpu(*vp++);
1299                 cipher_text.data = (__u8 *) vp;
1300                 vlen -= 4;
1301
1302                 if (vlen < cipher_text.len) {
1303                         CERROR("cipher text to be %u while buf only %u\n",
1304                                 cipher_text.len, vlen);
1305                         RETURN(-EPROTO);
1306                 }
1307
1308                 plain_text = cipher_text;
1309
1310                 gcred = container_of(cred, struct gss_cred, gc_base);
1311                 ctx = gss_cred_get_ctx(cred);
1312                 LASSERT(ctx);
1313
1314                 major = kgss_unwrap(ctx->gc_gss_ctx, GSS_C_QOP_DEFAULT,
1315                                     &cipher_text, &plain_text);
1316                 if (major) {
1317                         CERROR("cred %p: error unwrap: major 0x%x\n",
1318                                cred, major);
1319
1320                         if (major == GSS_S_CREDENTIALS_EXPIRED ||
1321                             major == GSS_S_CONTEXT_EXPIRED) {
1322                                 ptlrpcs_cred_expire(cred);
1323                                 req->rq_ptlrpcs_restart = 1;
1324                                 rc = 0;
1325                         } else
1326                                 rc = -EINVAL;
1327
1328                         GOTO(proc_out, rc);
1329                 }
1330
1331                 req->rq_repmsg = (struct lustre_msg *) vp;
1332                 req->rq_replen = plain_text.len;
1333
1334                 rc = 0;
1335 proc_out:
1336                 gss_put_ctx(ctx);
1337                 break;
1338         default:
1339                 CERROR("unknown gss proc %d\n", proc);
1340                 rc = -EPROTO;
1341         }
1342
1343         RETURN(rc);
1344 }
1345
1346 static void destroy_gss_context(struct ptlrpc_cred *cred)
1347 {
1348         struct ptlrpcs_wire_hdr *hdr;
1349         struct lustre_msg       *lmsg;
1350         struct gss_cred         *gcred;
1351         struct ptlrpc_request    req;
1352         struct obd_import       *imp;
1353         __u32                   *vp, lmsg_size;
1354         struct ptlrpc_request   *raw_req = NULL;
1355         const int                repbuf_len = 256;
1356         char                    *repbuf;
1357         int                      replen, rc;
1358         ENTRY;
1359
1360         imp = cred->pc_sec->ps_import;
1361         LASSERT(imp);
1362
1363         if (test_bit(PTLRPC_CRED_ERROR_BIT, &cred->pc_flags) ||
1364             !test_bit(PTLRPC_CRED_UPTODATE_BIT, &cred->pc_flags)) {
1365                 CDEBUG(D_SEC, "Destroy dead cred %p(%u@%s)\n",
1366                        cred, cred->pc_uid, imp->imp_target_uuid.uuid);
1367                 EXIT;
1368                 return;
1369         }
1370
1371         might_sleep();
1372
1373         /* cred's refcount is 0, steal one */
1374         atomic_inc(&cred->pc_refcount);
1375
1376         gcred = container_of(cred, struct gss_cred, gc_base);
1377         gcred->gc_ctx->gc_proc = PTLRPCS_GSS_PROC_DESTROY;
1378
1379         CDEBUG(D_SEC, "client destroy gss cred %p(%u@%s)\n",
1380                gcred, cred->pc_uid, imp->imp_target_uuid.uuid);
1381
1382         lmsg_size = lustre_msg_size(0, NULL);
1383         req.rq_req_secflvr = cred->pc_sec->ps_flavor;
1384         req.rq_cred = cred;
1385         req.rq_reqbuf_len = sizeof(*hdr) + lmsg_size +
1386                             ptlrpcs_est_req_payload(&req, lmsg_size);
1387
1388         OBD_ALLOC(req.rq_reqbuf, req.rq_reqbuf_len);
1389         if (!req.rq_reqbuf) {
1390                 CERROR("Fail to alloc reqbuf, cancel anyway\n");
1391                 atomic_dec(&cred->pc_refcount);
1392                 EXIT;
1393                 return;
1394         }
1395
1396         /* wire hdr */
1397         hdr = buf_to_sec_hdr(req.rq_reqbuf);
1398         hdr->flavor  = cpu_to_le32(PTLRPCS_FLVR_GSS_AUTH);
1399         hdr->msg_len = cpu_to_le32(lmsg_size);
1400         hdr->sec_len = cpu_to_le32(0);
1401
1402         /* lustre message */
1403         lmsg = buf_to_lustre_msg(req.rq_reqbuf);
1404         lustre_init_msg(lmsg, 0, NULL, NULL);
1405         lmsg->handle   = imp->imp_remote_handle;
1406         lmsg->type     = PTL_RPC_MSG_REQUEST;
1407         lmsg->opc      = SEC_FINI;
1408         lmsg->flags    = 0;
1409         lmsg->conn_cnt = imp->imp_conn_cnt;
1410         /* add this for randomize */
1411         get_random_bytes(&lmsg->last_xid, sizeof(lmsg->last_xid));
1412         get_random_bytes(&lmsg->transno, sizeof(lmsg->transno));
1413
1414         vp = (__u32 *) req.rq_reqbuf;
1415
1416         req.rq_cred = cred;
1417         req.rq_reqmsg = buf_to_lustre_msg(req.rq_reqbuf);
1418         req.rq_reqlen = lmsg_size;
1419         req.rq_reqdata_len = sizeof(*hdr) + lmsg_size;
1420
1421         if (gss_cred_sign(cred, &req)) {
1422                 CERROR("failed to sign, cancel anyway\n");
1423                 atomic_dec(&cred->pc_refcount);
1424                 goto exit;
1425         }
1426         atomic_dec(&cred->pc_refcount);
1427
1428         OBD_ALLOC(repbuf, repbuf_len);
1429         if (!repbuf)
1430                 goto exit;
1431
1432         raw_req = ptl_do_rawrpc(imp, req.rq_reqbuf, req.rq_reqbuf_len,
1433                                 req.rq_reqdata_len, repbuf, repbuf_len, &replen,
1434                                 SECFINI_RPC_TIMEOUT, &rc);
1435         if (!raw_req)
1436                 OBD_FREE(repbuf, repbuf_len);
1437
1438 exit:
1439         if (raw_req == NULL)
1440                 OBD_FREE(req.rq_reqbuf, req.rq_reqbuf_len);
1441         else
1442                 rawrpc_req_finished(raw_req);
1443         EXIT;
1444 }
1445
1446 static void gss_cred_destroy(struct ptlrpc_cred *cred)
1447 {
1448         struct gss_cred *gcred;
1449         ENTRY;
1450
1451         LASSERT(cred);
1452         LASSERT(!atomic_read(&cred->pc_refcount));
1453
1454         gcred = container_of(cred, struct gss_cred, gc_base);
1455         if (gcred->gc_ctx) {
1456                 destroy_gss_context(cred);
1457                 gss_put_ctx(gcred->gc_ctx);
1458         }
1459
1460         CDEBUG(D_SEC, "sec.gss %p: destroy cred %p\n", cred->pc_sec, gcred);
1461
1462         OBD_FREE(gcred, sizeof(*gcred));
1463         EXIT;
1464 }
1465
1466 static struct ptlrpc_credops gss_credops = {
1467         .refresh        = gss_cred_refresh,
1468         .match          = gss_cred_match,
1469         .sign           = gss_cred_sign,
1470         .verify         = gss_cred_verify,
1471         .seal           = gss_cred_seal,
1472         .unseal         = gss_cred_unseal,
1473         .destroy        = gss_cred_destroy,
1474 };
1475
1476 #ifdef __KERNEL__
1477 /*******************************************
1478  * rpc_pipe APIs                           *
1479  *******************************************/
1480 static ssize_t
1481 gss_pipe_upcall(struct file *filp, struct rpc_pipe_msg *msg,
1482                 char *dst, size_t buflen)
1483 {
1484         char *data = (char *)msg->data + msg->copied;
1485         ssize_t mlen = msg->len;
1486         ssize_t left;
1487         ENTRY;
1488
1489         if (mlen > buflen)
1490                 mlen = buflen;
1491         left = copy_to_user(dst, data, mlen);
1492         if (left < 0) {
1493                 msg->errno = left;
1494                 RETURN(left);
1495         }
1496         mlen -= left;
1497         msg->copied += mlen;
1498         msg->errno = 0;
1499         RETURN(mlen);
1500 }
1501
1502 static ssize_t
1503 gss_pipe_downcall(struct file *filp, const char *src, size_t mlen)
1504 {
1505         char *buf;
1506         const int bufsize = 1024;
1507         rawobj_t obj;
1508         struct inode *inode = filp->f_dentry->d_inode;
1509         struct rpc_inode *rpci = RPC_I(inode);
1510         struct obd_import *import;
1511         struct ptlrpc_sec *sec;
1512         struct gss_sec *gsec;
1513         char *obdname;
1514         struct gss_api_mech *mech;
1515         struct vfs_cred vcred = { 0 };
1516         struct ptlrpc_cred *cred;
1517         struct gss_upcall_msg *gss_msg;
1518         struct gss_upcall_msg_data gmd = { 0 };
1519         struct gss_cl_ctx *ctx = NULL;
1520         ssize_t left;
1521         int err, gss_err;
1522         ENTRY;
1523
1524         if (mlen > bufsize) {
1525                 CERROR("mlen %ld > bufsize %d\n", (long)mlen, bufsize);
1526                 RETURN(-ENOSPC);
1527         }
1528
1529         OBD_ALLOC(buf, bufsize);
1530         if (!buf) {
1531                 CERROR("alloc mem failed\n");
1532                 RETURN(-ENOMEM);
1533         }
1534
1535         left = copy_from_user(buf, src, mlen);
1536         if (left)
1537                 GOTO(err_free, err = -EFAULT);
1538
1539         obj.data = (unsigned char *)buf;
1540         obj.len = mlen;
1541
1542         LASSERT(rpci->private);
1543         gsec = (struct gss_sec *)rpci->private;
1544         sec = &gsec->gs_base;
1545         LASSERT(sec->ps_import);
1546         import = class_import_get(sec->ps_import);
1547         LASSERT(import->imp_obd);
1548         obdname = import->imp_obd->obd_name;
1549         mech = gsec->gs_mech;
1550
1551         err = gss_parse_init_downcall(mech, &obj, &ctx, &gmd, &gss_err);
1552         if (err)
1553                 CERROR("parse init downcall err %d\n", err);
1554
1555         vcred.vc_uid = gmd.gum_uid;
1556         vcred.vc_pag = vcred.vc_uid; /* FIXME */
1557
1558         cred = ptlrpcs_cred_lookup(sec, &vcred);
1559         if (!cred) {
1560                 CWARN("didn't find cred for uid %u\n", vcred.vc_uid);
1561                 GOTO(err, err = -EINVAL);
1562         }
1563
1564         if (err || gss_err) {
1565                 set_bit(PTLRPC_CRED_DEAD_BIT, &cred->pc_flags);
1566                 if (err != -ERESTART || gss_err != 0)
1567                         set_bit(PTLRPC_CRED_ERROR_BIT, &cred->pc_flags);
1568                 CERROR("cred %p: rpc err %d, gss err 0x%x, fatal %d\n",
1569                        cred, err, gss_err,
1570                        (test_bit(PTLRPC_CRED_ERROR_BIT, &cred->pc_flags) != 0));
1571         } else {
1572                 CDEBUG(D_SEC, "get initial ctx:\n");
1573                 gss_cred_set_ctx(cred, ctx);
1574         }
1575
1576         spin_lock(&gsec->gs_lock);
1577         gss_msg = gss_find_upcall(gsec, obdname, &gmd);
1578         if (gss_msg) {
1579                 gss_unhash_msg_nolock(gss_msg);
1580                 spin_unlock(&gsec->gs_lock);
1581                 gss_release_msg(gss_msg);
1582         } else
1583                 spin_unlock(&gsec->gs_lock);
1584
1585         ptlrpcs_cred_put(cred, 1);
1586         class_import_put(import);
1587         OBD_FREE(buf, bufsize);
1588         RETURN(mlen);
1589 err:
1590         if (ctx)
1591                 gss_destroy_ctx(ctx);
1592         class_import_put(import);
1593 err_free:
1594         OBD_FREE(buf, bufsize);
1595         CDEBUG(D_SEC, "gss_pipe_downcall returning %d\n", err);
1596         RETURN(err);
1597 }
1598
1599 static
1600 void gss_pipe_destroy_msg(struct rpc_pipe_msg *msg)
1601 {
1602         struct gss_upcall_msg *gmsg;
1603         static unsigned long ratelimit;
1604         ENTRY;
1605
1606         LASSERT(list_empty(&msg->list));
1607
1608         if (msg->errno >= 0) {
1609                 EXIT;
1610                 return;
1611         }
1612
1613         gmsg = container_of(msg, struct gss_upcall_msg, gum_base);
1614         CDEBUG(D_SEC, "destroy gmsg %p\n", gmsg);
1615         LASSERT(atomic_read(&gmsg->gum_refcount) > 0);
1616         atomic_inc(&gmsg->gum_refcount);
1617         gss_unhash_msg(gmsg);
1618         if (msg->errno == -ETIMEDOUT || msg->errno == -EPIPE) {
1619                 unsigned long now = get_seconds();
1620                 if (time_after(now, ratelimit)) {
1621                         CWARN("sec.gss upcall timed out.\n"
1622                               "Please check user daemon is running!\n");
1623                         ratelimit = now + 15;
1624                 }
1625         }
1626         gss_release_msg(gmsg);
1627         EXIT;
1628 }
1629
1630 static
1631 void gss_pipe_release(struct inode *inode)
1632 {
1633         struct rpc_inode *rpci = RPC_I(inode);
1634         struct ptlrpc_sec *sec;
1635         struct gss_sec *gsec;
1636         ENTRY;
1637
1638         gsec = (struct gss_sec *)rpci->private;
1639         sec = &gsec->gs_base;
1640         spin_lock(&gsec->gs_lock);
1641         while (!list_empty(&gsec->gs_upcalls)) {
1642                 struct gss_upcall_msg *gmsg;
1643
1644                 gmsg = list_entry(gsec->gs_upcalls.next,
1645                                   struct gss_upcall_msg, gum_list);
1646                 LASSERT(list_empty(&gmsg->gum_base.list));
1647                 gmsg->gum_base.errno = -EPIPE;
1648                 atomic_inc(&gmsg->gum_refcount);
1649                 gss_unhash_msg_nolock(gmsg);
1650                 gss_release_msg(gmsg);
1651         }
1652         spin_unlock(&gsec->gs_lock);
1653         EXIT;
1654 }
1655
1656 static struct rpc_pipe_ops gss_upcall_ops = {
1657         .upcall         = gss_pipe_upcall,
1658         .downcall       = gss_pipe_downcall,
1659         .destroy_msg    = gss_pipe_destroy_msg,
1660         .release_pipe   = gss_pipe_release,
1661 };
1662 #endif /* __KERNEL__ */
1663
1664 /*********************************************
1665  * GSS security APIs                         *
1666  *********************************************/
1667
1668 static
1669 struct ptlrpc_sec* gss_create_sec(__u32 flavor,
1670                                   const char *pipe_dir,
1671                                   void *pipe_data)
1672 {
1673         struct gss_sec *gsec;
1674         struct ptlrpc_sec *sec;
1675 #ifdef __KERNEL__
1676         char *pos;
1677         int   pipepath_len;
1678         uid_t save_uid;
1679 #endif
1680         ENTRY;
1681
1682         LASSERT(SEC_FLAVOR_MAJOR(flavor) == PTLRPCS_FLVR_MAJOR_GSS);
1683
1684         OBD_ALLOC(gsec, sizeof(*gsec));
1685         if (!gsec) {
1686                 CERROR("can't alloc gsec\n");
1687                 RETURN(NULL);
1688         }
1689
1690         gsec->gs_mech = kgss_subflavor_to_mech(SEC_FLAVOR_SUB(flavor));
1691         if (!gsec->gs_mech) {
1692                 CERROR("subflavor 0x%x not found\n", flavor);
1693                 goto err_free;
1694         }
1695
1696         /* initialize gss sec */
1697 #ifdef __KERNEL__
1698         INIT_LIST_HEAD(&gsec->gs_upcalls);
1699         spin_lock_init(&gsec->gs_lock);
1700
1701         pipepath_len = strlen(LUSTRE_PIPEDIR) + strlen(pipe_dir) +
1702                        strlen(gsec->gs_mech->gm_name) + 3;
1703         OBD_ALLOC(gsec->gs_pipepath, pipepath_len);
1704         if (!gsec->gs_pipepath)
1705                 goto err_mech_put;
1706
1707         /* pipe rpc require root permission */
1708         save_uid = current->fsuid;
1709         current->fsuid = 0;
1710
1711         sprintf(gsec->gs_pipepath, LUSTRE_PIPEDIR"/%s", pipe_dir);
1712         if (IS_ERR(rpc_mkdir(gsec->gs_pipepath, NULL))) {
1713                 CERROR("can't make pipedir %s\n", gsec->gs_pipepath);
1714                 goto err_free_path;
1715         }
1716
1717         sprintf(gsec->gs_pipepath, LUSTRE_PIPEDIR"/%s/%s", pipe_dir,
1718                 gsec->gs_mech->gm_name);
1719         gsec->gs_depipe = rpc_mkpipe(gsec->gs_pipepath, gsec,
1720                                      &gss_upcall_ops, RPC_PIPE_WAIT_FOR_OPEN);
1721         if (IS_ERR(gsec->gs_depipe)) {
1722                 CERROR("failed to make rpc_pipe %s: %ld\n",
1723                         gsec->gs_pipepath, PTR_ERR(gsec->gs_depipe));
1724                 goto err_rmdir;
1725         }
1726         CDEBUG(D_SEC, "gss sec %p, pipe path %s\n", gsec, gsec->gs_pipepath);
1727 #endif
1728
1729         sec = &gsec->gs_base;
1730         sec->ps_expire = GSS_CREDCACHE_EXPIRE;
1731         sec->ps_nextgc = get_seconds() + sec->ps_expire;
1732         sec->ps_flags = 0;
1733
1734         current->fsuid = save_uid;
1735
1736         CDEBUG(D_SEC, "Create sec.gss %p\n", gsec);
1737         RETURN(sec);
1738
1739 #ifdef __KERNEL__
1740 err_rmdir:
1741         pos = strrchr(gsec->gs_pipepath, '/');
1742         LASSERT(pos);
1743         *pos = 0;
1744         rpc_rmdir(gsec->gs_pipepath);
1745 err_free_path:
1746         current->fsuid = save_uid;
1747         OBD_FREE(gsec->gs_pipepath, pipepath_len);
1748 err_mech_put:
1749 #endif
1750         kgss_mech_put(gsec->gs_mech);
1751 err_free:
1752         OBD_FREE(gsec, sizeof(*gsec));
1753         RETURN(NULL);
1754 }
1755
1756 static
1757 void gss_destroy_sec(struct ptlrpc_sec *sec)
1758 {
1759         struct gss_sec *gsec;
1760 #ifdef __KERNEL__
1761         char *pos;
1762         int   pipepath_len;
1763 #endif
1764         ENTRY;
1765
1766         gsec = container_of(sec, struct gss_sec, gs_base);
1767         CDEBUG(D_SEC, "Destroy sec.gss %p\n", gsec);
1768
1769         LASSERT(gsec->gs_mech);
1770         LASSERT(!atomic_read(&sec->ps_refcount));
1771         LASSERT(!atomic_read(&sec->ps_credcount));
1772 #ifdef __KERNEL__
1773         pipepath_len = strlen(gsec->gs_pipepath) + 1;
1774         rpc_unlink(gsec->gs_pipepath);
1775         pos = strrchr(gsec->gs_pipepath, '/');
1776         LASSERT(pos);
1777         *pos = 0;
1778         rpc_rmdir(gsec->gs_pipepath);
1779         OBD_FREE(gsec->gs_pipepath, pipepath_len);
1780 #endif
1781
1782         kgss_mech_put(gsec->gs_mech);
1783         OBD_FREE(gsec, sizeof(*gsec));
1784         EXIT;
1785 }
1786
1787 static
1788 struct ptlrpc_cred * gss_create_cred(struct ptlrpc_sec *sec,
1789                                      struct vfs_cred *vcred)
1790 {
1791         struct gss_cred *gcred;
1792         struct ptlrpc_cred *cred;
1793         ENTRY;
1794
1795         OBD_ALLOC(gcred, sizeof(*gcred));
1796         if (!gcred)
1797                 RETURN(NULL);
1798
1799         cred = &gcred->gc_base;
1800         INIT_LIST_HEAD(&cred->pc_hash);
1801         atomic_set(&cred->pc_refcount, 0);
1802         cred->pc_sec = sec;
1803         cred->pc_ops = &gss_credops;
1804         cred->pc_expire = 0;
1805         cred->pc_flags = 0;
1806         cred->pc_pag = vcred->vc_pag;
1807         cred->pc_uid = vcred->vc_uid;
1808         CDEBUG(D_SEC, "create a gss cred at %p("LPU64"/%u)\n",
1809                cred, vcred->vc_pag, vcred->vc_uid);
1810
1811         RETURN(cred);
1812 }
1813
1814 static int gss_estimate_payload(struct ptlrpc_sec *sec,
1815                                 struct ptlrpc_request *req,
1816                                 int msgsize)
1817 {
1818         switch (SEC_FLAVOR_SVC(req->rq_req_secflvr)) {
1819         case PTLRPCS_SVC_AUTH:
1820                 return GSS_MAX_AUTH_PAYLOAD;
1821         case PTLRPCS_SVC_PRIV:
1822                 return size_round16(GSS_MAX_AUTH_PAYLOAD + msgsize +
1823                                     GSS_PRIVBUF_PREFIX_LEN +
1824                                     GSS_PRIVBUF_SUFFIX_LEN);
1825         default:
1826                 LBUG();
1827                 return 0;
1828         }
1829 }
1830
1831 static int gss_alloc_reqbuf(struct ptlrpc_sec *sec,
1832                             struct ptlrpc_request *req,
1833                             int lmsg_size)
1834 {
1835         int msg_payload, sec_payload;
1836         int privacy, rc;
1837         ENTRY;
1838
1839         /* In PRIVACY mode, lustre message is always 0 (already encoded into
1840          * security payload).
1841          */
1842         privacy = (SEC_FLAVOR_SVC(req->rq_req_secflvr) == PTLRPCS_SVC_PRIV);
1843         msg_payload = privacy ? 0 : lmsg_size;
1844         sec_payload = gss_estimate_payload(sec, req, lmsg_size);
1845
1846         rc = sec_alloc_reqbuf(sec, req, msg_payload, sec_payload);
1847         if (rc)
1848                 return rc;
1849
1850         if (privacy) {
1851                 int buflen = lmsg_size + GSS_PRIVBUF_PREFIX_LEN +
1852                              GSS_PRIVBUF_SUFFIX_LEN;
1853                 char *buf;
1854
1855                 OBD_ALLOC(buf, buflen);
1856                 if (!buf) {
1857                         CERROR("Fail to alloc %d\n", buflen);
1858                         sec_free_reqbuf(sec, req);
1859                         RETURN(-ENOMEM);
1860                 }
1861                 req->rq_reqmsg = (struct lustre_msg *)
1862                                         (buf + GSS_PRIVBUF_PREFIX_LEN);
1863         }
1864
1865         RETURN(0);
1866 }
1867
1868 static void gss_free_reqbuf(struct ptlrpc_sec *sec,
1869                             struct ptlrpc_request *req)
1870 {
1871         char *buf;
1872         int privacy;
1873         ENTRY;
1874
1875         LASSERT(req->rq_reqmsg);
1876         LASSERT(req->rq_reqlen);
1877
1878         privacy = SEC_FLAVOR_SVC(req->rq_req_secflvr) == PTLRPCS_SVC_PRIV;
1879         if (privacy) {
1880                 buf = (char *) req->rq_reqmsg - GSS_PRIVBUF_PREFIX_LEN;
1881                 LASSERT(buf < req->rq_reqbuf ||
1882                         buf >= req->rq_reqbuf + req->rq_reqbuf_len);
1883                 OBD_FREE(buf, req->rq_reqlen + GSS_PRIVBUF_PREFIX_LEN +
1884                               GSS_PRIVBUF_SUFFIX_LEN);
1885                 req->rq_reqmsg = NULL;
1886         }
1887
1888         sec_free_reqbuf(sec, req);
1889 }
1890
1891 static struct ptlrpc_secops gss_secops = {
1892         .create_sec             = gss_create_sec,
1893         .destroy_sec            = gss_destroy_sec,
1894         .create_cred            = gss_create_cred,
1895         .est_req_payload        = gss_estimate_payload,
1896         .est_rep_payload        = gss_estimate_payload,
1897         .alloc_reqbuf           = gss_alloc_reqbuf,
1898         .free_reqbuf            = gss_free_reqbuf,
1899 };
1900
1901 static struct ptlrpc_sec_type gss_type = {
1902         .pst_owner      = THIS_MODULE,
1903         .pst_name       = "sec.gss",
1904         .pst_inst       = ATOMIC_INIT(0),
1905         .pst_flavor     = PTLRPCS_FLVR_MAJOR_GSS,
1906         .pst_ops        = &gss_secops,
1907 };
1908
1909 extern int
1910 (*lustre_secinit_downcall_handler)(char *buffer, unsigned long count);
1911
1912 int __init ptlrpcs_gss_init(void)
1913 {
1914         int rc;
1915
1916         rc = ptlrpcs_register(&gss_type);
1917         if (rc)
1918                 return rc;
1919
1920 #ifdef __KERNEL__
1921         gss_svc_init();
1922
1923         rc = PTR_ERR(rpc_mkdir(LUSTRE_PIPEDIR, NULL));
1924         if (IS_ERR((void *)rc) && rc != -EEXIST) {
1925                 CERROR("fail to make rpcpipedir for lustre\n");
1926                 gss_svc_exit();
1927                 ptlrpcs_unregister(&gss_type);
1928                 return -1;
1929         }
1930         rc = 0;
1931 #else
1932 #endif
1933         rc = init_kerberos_module();
1934         if (rc) {
1935                 ptlrpcs_unregister(&gss_type);
1936         }
1937
1938         lustre_secinit_downcall_handler = gss_send_secinit_rpc;
1939
1940         return rc;
1941 }
1942
1943 #ifdef __KERNEL__
1944 static void __exit ptlrpcs_gss_exit(void)
1945 {
1946         lustre_secinit_downcall_handler = NULL;
1947
1948         cleanup_kerberos_module();
1949         rpc_rmdir(LUSTRE_PIPEDIR);
1950         gss_svc_exit();
1951         ptlrpcs_unregister(&gss_type);
1952 }
1953 #endif
1954
1955 MODULE_AUTHOR("Cluster File Systems, Inc. <info@clusterfs.com>");
1956 MODULE_DESCRIPTION("GSS Security module for Lustre");
1957 MODULE_LICENSE("GPL");
1958
1959 module_init(ptlrpcs_gss_init);
1960 module_exit(ptlrpcs_gss_exit);