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         spinlock_t              gs_lock;
 388         struct list_head        gs_upcalls;
 389         char                   *gs_pipepath;
 390         struct dentry          *gs_depipe;
 391 };
 392
 393 struct gss_upcall_msg_data {
 394         __u64                           gum_pag;
 395         __u32                           gum_uid;
 396         __u32                           gum_svc;
 397         __u32                           gum_nal;
 398         __u32                           gum_netid;
 399         __u64                           gum_nid;
 400 };
 401
 402 struct gss_upcall_msg {
 403         struct rpc_pipe_msg             gum_base;
 404         atomic_t                        gum_refcount;
 405         struct list_head                gum_list;
 406         struct gss_sec                 *gum_gsec;
 407         wait_queue_head_t               gum_waitq;
 408         char                            gum_obdname[64];
 409         struct gss_upcall_msg_data      gum_data;
 410 };
 411
 412 #ifdef __KERNEL__
 413 static rwlock_t gss_ctx_lock = RW_LOCK_UNLOCKED;
 414 /**********************************************
 415  * rpc_pipe upcall helpers                    *
 416  **********************************************/
 417 static
 418 void gss_release_msg(struct gss_upcall_msg *gmsg)
 419 {
 420         ENTRY;
 421         LASSERT(atomic_read(&gmsg->gum_refcount) > 0);
 422
 423         if (!atomic_dec_and_test(&gmsg->gum_refcount)) {
 424                 CDEBUG(D_SEC, "gmsg %p ref %d\n", gmsg,
 425                        atomic_read(&gmsg->gum_refcount));
 426                 EXIT;
 427                 return;
 428         }
 429         LASSERT(list_empty(&gmsg->gum_list));
 430 #if 0
 431         LASSERT(list_empty(&gmsg->gum_base.list));
 432 #else
 433         /* XXX */
 434         if (!list_empty(&gmsg->gum_base.list)) {
 435                 int error = gmsg->gum_base.errno;
 436
 437                 CWARN("msg %p: list: %p/%p/%p, copied %d, err %d, wq %d\n",
 438                       gmsg, &gmsg->gum_base.list, gmsg->gum_base.list.prev,
 439                       gmsg->gum_base.list.next, gmsg->gum_base.copied, error,
 440                       list_empty(&gmsg->gum_waitq.task_list));
 441                 LBUG();
 442         }
 443 #endif
 444         OBD_FREE(gmsg, sizeof(*gmsg));
 445         EXIT;
 446 }
 447
 448 static void
 449 gss_unhash_msg_nolock(struct gss_upcall_msg *gmsg)
 450 {
 451         LASSERT_SPIN_LOCKED(&gmsg->gum_gsec->gs_lock);
 452
 453         if (list_empty(&gmsg->gum_list))
 454                 return;
 455
 456         list_del_init(&gmsg->gum_list);
 457         wake_up(&gmsg->gum_waitq);
 458         LASSERT(atomic_read(&gmsg->gum_refcount) > 1);
 459         atomic_dec(&gmsg->gum_refcount);
 460 }
 461
 462 static void
 463 gss_unhash_msg(struct gss_upcall_msg *gmsg)
 464 {
 465         struct gss_sec *gsec = gmsg->gum_gsec;
 466
 467         spin_lock(&gsec->gs_lock);
 468         gss_unhash_msg_nolock(gmsg);
 469         spin_unlock(&gsec->gs_lock);
 470 }
 471
 472 static
 473 struct gss_upcall_msg * gss_find_upcall(struct gss_sec *gsec,
 474                                         char *obdname,
 475                                         struct gss_upcall_msg_data *gmd)
 476 {
 477         struct gss_upcall_msg *gmsg;
 478         ENTRY;
 479
 480         LASSERT_SPIN_LOCKED(&gsec->gs_lock);
 481
 482         list_for_each_entry(gmsg, &gsec->gs_upcalls, gum_list) {
 483                 if (memcmp(&gmsg->gum_data, gmd, sizeof(*gmd)))
 484                         continue;
 485                 if (strcmp(gmsg->gum_obdname, obdname))
 486                         continue;
 487                 LASSERT(atomic_read(&gmsg->gum_refcount) > 0);
 488                 atomic_inc(&gmsg->gum_refcount);
 489                 CDEBUG(D_SEC, "found gmsg at %p: obdname %s, uid %d, ref %d\n",
 490                        gmsg, obdname, gmd->gum_uid,
 491                        atomic_read(&gmsg->gum_refcount));
 492                 RETURN(gmsg);
 493         }
 494         RETURN(NULL);
 495 }
 496
 497 static void gss_init_upcall_msg(struct gss_upcall_msg *gmsg,
 498                                 struct gss_sec *gsec, char *obdname,
 499                                 struct gss_upcall_msg_data *gmd)
 500 {
 501         struct rpc_pipe_msg *rpcmsg;
 502         ENTRY;
 503
 504         /* 2 refs: 1 for hash, 1 for current user */
 505         init_waitqueue_head(&gmsg->gum_waitq);
 506         list_add(&gmsg->gum_list, &gsec->gs_upcalls);
 507         atomic_set(&gmsg->gum_refcount, 2);
 508         gmsg->gum_gsec = gsec;
 509         strncpy(gmsg->gum_obdname, obdname, sizeof(gmsg->gum_obdname));
 510         memcpy(&gmsg->gum_data, gmd, sizeof(*gmd));
 511
 512         rpcmsg = &gmsg->gum_base;
 513         INIT_LIST_HEAD(&rpcmsg->list);
 514         rpcmsg->data = &gmsg->gum_data;
 515         rpcmsg->len = sizeof(gmsg->gum_data);
 516         rpcmsg->copied = 0;
 517         rpcmsg->errno = 0;
 518         EXIT;
 519 }
 520 #endif /* __KERNEL__ */
 521
 522 /* this seems to be used only from userspace code */
 523 #ifndef __KERNEL__
 524 /********************************************
 525  * gss cred manipulation helpers            *
 526  ********************************************/
 527 static
 528 int gss_cred_is_uptodate_ctx(struct ptlrpc_cred *cred)
 529 {
 530         struct gss_cred *gcred = container_of(cred, struct gss_cred, gc_base);
 531         int res = 0;
 532
 533         read_lock(&gss_ctx_lock);
 534         if (((cred->pc_flags & PTLRPC_CRED_FLAGS_MASK) ==
 535              PTLRPC_CRED_UPTODATE) &&
 536             gcred->gc_ctx)
 537                 res = 1;
 538         read_unlock(&gss_ctx_lock);
 539         return res;
 540 }
 541 #endif
 542
 543 static inline
 544 struct gss_cl_ctx *gss_get_ctx(struct gss_cl_ctx *ctx)
 545 {
 546         atomic_inc(&ctx->gc_refcount);
 547         return ctx;
 548 }
 549
 550 static
 551 void gss_destroy_ctx(struct gss_cl_ctx *ctx)
 552 {
 553         ENTRY;
 554
 555         CDEBUG(D_SEC, "destroy cl_ctx %p\n", ctx);
 556         if (ctx->gc_gss_ctx)
 557                 kgss_delete_sec_context(&ctx->gc_gss_ctx);
 558
 559         if (ctx->gc_wire_ctx.len > 0) {
 560                 OBD_FREE(ctx->gc_wire_ctx.data, ctx->gc_wire_ctx.len);
 561                 ctx->gc_wire_ctx.len = 0;
 562         }
 563
 564         OBD_FREE(ctx, sizeof(*ctx));
 565 }
 566
 567 static
 568 void gss_put_ctx(struct gss_cl_ctx *ctx)
 569 {
 570         if (atomic_dec_and_test(&ctx->gc_refcount))
 571                 gss_destroy_ctx(ctx);
 572 }
 573
 574 static
 575 struct gss_cl_ctx *gss_cred_get_ctx(struct ptlrpc_cred *cred)
 576 {
 577         struct gss_cred *gcred = container_of(cred, struct gss_cred, gc_base);
 578         struct gss_cl_ctx *ctx = NULL;
 579
 580         read_lock(&gss_ctx_lock);
 581         if (gcred->gc_ctx)
 582                 ctx = gss_get_ctx(gcred->gc_ctx);
 583         read_unlock(&gss_ctx_lock);
 584         return ctx;
 585 }
 586
 587 static
 588 void gss_cred_set_ctx(struct ptlrpc_cred *cred, struct gss_cl_ctx *ctx)
 589 {
 590         struct gss_cred *gcred = container_of(cred, struct gss_cred, gc_base);
 591         struct gss_cl_ctx *old;
 592         __u64 ctx_expiry;
 593         ENTRY;
 594
 595         if (kgss_inquire_context(ctx->gc_gss_ctx, &ctx_expiry)) {
 596                 CERROR("unable to get expire time\n");
 597                 ctx_expiry = 1; /* make it expired now */
 598         }
 599         cred->pc_expire = (unsigned long) ctx_expiry;
 600
 601         write_lock(&gss_ctx_lock);
 602         old = gcred->gc_ctx;
 603         gcred->gc_ctx = ctx;
 604         set_bit(PTLRPC_CRED_UPTODATE_BIT, &cred->pc_flags);
 605         write_unlock(&gss_ctx_lock);
 606         if (old)
 607                 gss_put_ctx(old);
 608
 609         CDEBUG(D_SEC, "client refreshed gss cred %p(uid %u)\n",
 610                cred, cred->pc_uid);
 611         EXIT;
 612 }
 613
 614 static int
 615 simple_get_bytes(char **buf, __u32 *buflen, void *res, __u32 reslen)
 616 {
 617         if (*buflen < reslen) {
 618                 CERROR("buflen %u < %u\n", *buflen, reslen);
 619                 return -EINVAL;
 620         }
 621
 622         memcpy(res, *buf, reslen);
 623         *buf += reslen;
 624         *buflen -= reslen;
 625         return 0;
 626 }
 627
 628 /* data passed down:
 629  *  - uid
 630  *  - timeout
 631  *  - gc_win / error
 632  *  - wire_ctx (rawobj)
 633  *  - mech_ctx? (rawobj)
 634  */
 635 static
 636 int gss_parse_init_downcall(struct gss_api_mech *gm, rawobj_t *buf,
 637                             struct gss_cl_ctx **gc,
 638                             struct gss_upcall_msg_data *gmd, int *gss_err)
 639 {
 640         char *p = (char *)buf->data;
 641         struct gss_cl_ctx *ctx;
 642         __u32 len = buf->len;
 643         unsigned int timeout;
 644         rawobj_t tmp_buf;
 645         int err = -EPERM;
 646         ENTRY;
 647
 648         *gc = NULL;
 649         *gss_err = 0;
 650
 651         OBD_ALLOC(ctx, sizeof(*ctx));
 652         if (!ctx)
 653                 RETURN(-ENOMEM);
 654
 655         ctx->gc_proc = RPC_GSS_PROC_DATA;
 656         ctx->gc_seq = 0;
 657         spin_lock_init(&ctx->gc_seq_lock);
 658         atomic_set(&ctx->gc_refcount,1);
 659
 660         if (simple_get_bytes(&p, &len, &gmd->gum_pag, sizeof(gmd->gum_pag)))
 661                 goto err_free_ctx;
 662         if (simple_get_bytes(&p, &len, &gmd->gum_uid, sizeof(gmd->gum_uid)))
 663                 goto err_free_ctx;
 664         if (simple_get_bytes(&p, &len, &gmd->gum_svc, sizeof(gmd->gum_svc)))
 665                 goto err_free_ctx;
 666         if (simple_get_bytes(&p, &len, &gmd->gum_nal, sizeof(gmd->gum_nal)))
 667                 goto err_free_ctx;
 668         if (simple_get_bytes(&p, &len, &gmd->gum_netid, sizeof(gmd->gum_netid)))
 669                 goto err_free_ctx;
 670         if (simple_get_bytes(&p, &len, &gmd->gum_nid, sizeof(gmd->gum_nid)))
 671                 goto err_free_ctx;
 672         /* FIXME: discarded timeout for now */
 673         if (simple_get_bytes(&p, &len, &timeout, sizeof(timeout)))
 674                 goto err_free_ctx;
 675         if (simple_get_bytes(&p, &len, &ctx->gc_win, sizeof(ctx->gc_win)))
 676                 goto err_free_ctx;
 677
 678         /* lgssd signals an error by passing ctx->gc_win = 0: */
 679         if (!ctx->gc_win) {
 680                 /* in which case the next 2 int are:
 681                  * - rpc error
 682                  * - gss error
 683                  */
 684                 if (simple_get_bytes(&p, &len, &err, sizeof(err))) {
 685                         err = -EPERM;
 686                         goto err_free_ctx;
 687                 }
 688                 if (simple_get_bytes(&p, &len, gss_err, sizeof(*gss_err))) {
 689                         err = -EPERM;
 690                         goto err_free_ctx;
 691                 }
 692                 if (err == 0 && *gss_err == 0) {
 693                         CERROR("no error passed from downcall\n");
 694                         err = -EPERM;
 695                 }
 696                 goto err_free_ctx;
 697         }
 698
 699         if (rawobj_extract_local(&tmp_buf, (__u32 **) ((void *)&p), &len))
 700                 goto err_free_ctx;
 701         if (rawobj_dup(&ctx->gc_wire_ctx, &tmp_buf)) {
 702                 err = -ENOMEM;
 703                 goto err_free_ctx;
 704         }
 705         if (rawobj_extract_local(&tmp_buf, (__u32 **) ((void *)&p), &len))
 706                 goto err_free_wire_ctx;
 707         if (len) {
 708                 CERROR("unexpected trailing %u bytes\n", len);
 709                 goto err_free_wire_ctx;
 710         }
 711         if (kgss_import_sec_context(&tmp_buf, gm, &ctx->gc_gss_ctx))
 712                 goto err_free_wire_ctx;
 713
 714         *gc = ctx;
 715         RETURN(0);
 716
 717 err_free_wire_ctx:
 718         if (ctx->gc_wire_ctx.data)
 719                 OBD_FREE(ctx->gc_wire_ctx.data, ctx->gc_wire_ctx.len);
 720 err_free_ctx:
 721         OBD_FREE(ctx, sizeof(*ctx));
 722         CDEBUG(D_SEC, "err_code %d, gss code %d\n", err, *gss_err);
 723         return err;
 724 }
 725
 726 /***************************************
 727  * cred APIs                           *
 728  ***************************************/
 729 #ifdef __KERNEL__
 730 static int gss_cred_refresh(struct ptlrpc_cred *cred)
 731 {
 732         struct obd_import          *import;
 733         struct gss_sec             *gsec;
 734         struct gss_upcall_msg      *gss_msg, *gss_new;
 735         struct gss_upcall_msg_data  gmd;
 736         struct dentry              *dentry;
 737         char                       *obdname, *obdtype;
 738         wait_queue_t                wait;
 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_pag = cred->pc_pag;
 759         gmd.gum_uid = cred->pc_uid;
 760         gmd.gum_nal = import->imp_connection->c_peer.peer_ni->pni_number;
 761         gmd.gum_netid = 0;
 762         gmd.gum_nid = import->imp_connection->c_peer.peer_id.nid;
 763
 764         obdtype = import->imp_obd->obd_type->typ_name;
 765         if (!strcmp(obdtype, OBD_MDC_DEVICENAME))
 766                 gmd.gum_svc = LUSTRE_GSS_SVC_MDS;
 767         else if (!strcmp(obdtype, OBD_OSC_DEVICENAME))
 768                 gmd.gum_svc = LUSTRE_GSS_SVC_OSS;
 769         else {
 770                 CERROR("gss on %s?\n", obdtype);
 771                 RETURN(-EINVAL);
 772         }
 773
 774         gsec = container_of(cred->pc_sec, struct gss_sec, gs_base);
 775         obdname = import->imp_obd->obd_name;
 776         dentry = gsec->gs_depipe;
 777         gss_new = NULL;
 778         res = 0;
 779
 780         CDEBUG(D_SEC, "Initiate gss context %p(%u@%s)\n",
 781                container_of(cred, struct gss_cred, gc_base),
 782                cred->pc_uid, import->imp_target_uuid.uuid);
 783
 784 again:
 785         spin_lock(&gsec->gs_lock);
 786         gss_msg = gss_find_upcall(gsec, obdname, &gmd);
 787         if (gss_msg) {
 788                 if (gss_new) {
 789                         OBD_FREE(gss_new, sizeof(*gss_new));
 790                         gss_new = NULL;
 791                 }
 792                 GOTO(waiting, res);
 793         }
 794
 795         if (!gss_new) {
 796                 spin_unlock(&gsec->gs_lock);
 797                 OBD_ALLOC(gss_new, sizeof(*gss_new));
 798                 if (!gss_new)
 799                         RETURN(-ENOMEM);
 800                 goto again;
 801         }
 802         /* so far we'v created gss_new */
 803         gss_init_upcall_msg(gss_new, gsec, obdname, &gmd);
 804
 805         /* we'v created upcall msg, nobody else should touch the
 806          * flag of this cred, unless be set as dead/expire by
 807          * administrator via lctl etc.
 808          */
 809         if (cred->pc_flags & PTLRPC_CRED_FLAGS_MASK) {
 810                 CWARN("cred %p("LPU64"/%u) was set flags %lx unexpectedly\n",
 811                       cred, cred->pc_pag, cred->pc_uid, cred->pc_flags);
 812                 cred->pc_flags |= PTLRPC_CRED_DEAD | PTLRPC_CRED_ERROR;
 813                 gss_unhash_msg_nolock(gss_new);
 814                 spin_unlock(&gsec->gs_lock);
 815                 gss_release_msg(gss_new);
 816                 RETURN(0);
 817         }
 818
 819         /* need to make upcall now */
 820         spin_unlock(&gsec->gs_lock);
 821         res = rpc_queue_upcall(dentry->d_inode, &gss_new->gum_base);
 822         if (res) {
 823                 CERROR("rpc_queue_upcall failed: %d\n", res);
 824                 gss_unhash_msg(gss_new);
 825                 gss_release_msg(gss_new);
 826                 cred->pc_flags |= PTLRPC_CRED_DEAD | PTLRPC_CRED_ERROR;
 827                 RETURN(res);
 828         }
 829         gss_msg = gss_new;
 830         spin_lock(&gsec->gs_lock);
 831
 832 waiting:
 833         /* upcall might finish quickly */
 834         if (list_empty(&gss_msg->gum_list)) {
 835                 spin_unlock(&gsec->gs_lock);
 836                 res = 0;
 837                 goto out;
 838         }
 839
 840         init_waitqueue_entry(&wait, current);
 841         set_current_state(TASK_INTERRUPTIBLE);
 842         add_wait_queue(&gss_msg->gum_waitq, &wait);
 843         spin_unlock(&gsec->gs_lock);
 844
 845         if (gss_new)
 846                 res = schedule_timeout(CRED_REFRESH_UPCALL_TIMEOUT * HZ);
 847         else {
 848                 schedule();
 849                 res = 0;
 850         }
 851
 852         remove_wait_queue(&gss_msg->gum_waitq, &wait);
 853
 854         /* - the one who refresh the cred for us should also be responsible
 855          *   to set the status of cred, we can simply return.
 856          * - if cred flags has been set, we also don't need to do that again,
 857          *   no matter signal pending or timeout etc.
 858          */
 859         if (!gss_new || cred->pc_flags & PTLRPC_CRED_FLAGS_MASK)
 860                 goto out;
 861
 862         if (signal_pending(current)) {
 863                 CERROR("%s: cred %p: interrupted upcall\n",
 864                        current->comm, cred);
 865                 cred->pc_flags |= PTLRPC_CRED_DEAD | PTLRPC_CRED_ERROR;
 866                 res = -EINTR;
 867         } else if (res == 0) {
 868                 CERROR("cred %p: upcall timedout\n", cred);
 869                 set_bit(PTLRPC_CRED_DEAD_BIT, &cred->pc_flags);
 870                 res = -ETIMEDOUT;
 871         } else
 872                 res = 0;
 873
 874 out:
 875         gss_release_msg(gss_msg);
 876
 877         RETURN(res);
 878 }
 879 #else /* !__KERNEL__ */
 880 extern int lgss_handle_krb5_upcall(uid_t uid, __u32 dest_ip,
 881                                    char *obd_name, 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         ptl_nid_t               peer_nid;
 893         __u32                   dest_ip;
 894         __u32                   subflavor;
 895         int                     rc, gss_err;
 896         struct gss_upcall_msg_data gmd = { 0 };
 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;
 910
 911         if (subflavor != PTLRPCS_SUBFLVR_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
 931         rc = gss_parse_init_downcall(mech, &obj, &ctx, &gmd,
 932                                      &gss_err);
 933         if (rc || gss_err) {
 934                 CERROR("parse init downcall: rpc %d, gss 0x%x\n", rc, gss_err);
 935                 if (rc != -ERESTART || gss_err != 0)
 936                         set_bit(PTLRPC_CRED_ERROR_BIT, &cred->pc_flags);
 937                 if (rc == 0)
 938                         rc = -EPERM;
 939                 goto err_out;
 940         }
 941
 942         LASSERT(ctx);
 943         gss_cred_set_ctx(cred, ctx);
 944         LASSERT(gss_cred_is_uptodate_ctx(cred));
 945
 946         return 0;
 947 err_out:
 948         set_bit(PTLRPC_CRED_DEAD_BIT, &cred->pc_flags);
 949         return rc;
 950 }
 951 #endif
 952
 953 static int gss_cred_match(struct ptlrpc_cred *cred,
 954                           struct vfs_cred *vcred)
 955 {
 956         RETURN(cred->pc_pag == vcred->vc_pag);
 957 }
 958
 959 static int gss_cred_sign(struct ptlrpc_cred *cred,
 960                          struct ptlrpc_request *req)
 961 {
 962         struct gss_cred         *gcred;
 963         struct gss_cl_ctx       *ctx;
 964         rawobj_t                 lmsg, mic;
 965         __u32                   *vp, *vpsave, vlen, seclen;
 966         __u32                    seqnum, major, rc = 0;
 967         ENTRY;
 968
 969         LASSERT(req->rq_reqbuf);
 970         LASSERT(req->rq_cred == cred);
 971
 972         gcred = container_of(cred, struct gss_cred, gc_base);
 973         ctx = gss_cred_get_ctx(cred);
 974         if (!ctx) {
 975                 CERROR("cred %p("LPU64"/%u) invalidated?\n",
 976                         cred, cred->pc_pag, cred->pc_uid);
 977                 RETURN(-EPERM);
 978         }
 979
 980         lmsg.len = req->rq_reqlen;
 981         lmsg.data = (__u8 *) req->rq_reqmsg;
 982
 983         vp = (__u32 *) (lmsg.data + lmsg.len);
 984         vlen = req->rq_reqbuf_len - sizeof(struct ptlrpcs_wire_hdr) -
 985                lmsg.len;
 986         seclen = vlen;
 987
 988         if (vlen < 6 * 4 + size_round4(ctx->gc_wire_ctx.len)) {
 989                 CERROR("vlen %d, need %d\n",
 990                         vlen, 6 * 4 + size_round4(ctx->gc_wire_ctx.len));
 991                 rc = -EIO;
 992                 goto out;
 993         }
 994
 995         spin_lock(&ctx->gc_seq_lock);
 996         seqnum = ctx->gc_seq++;
 997         spin_unlock(&ctx->gc_seq_lock);
 998
 999         *vp++ = cpu_to_le32(PTLRPC_SEC_GSS_VERSION);    /* version */
1000         *vp++ = cpu_to_le32(PTLRPCS_FLVR_KRB5I);        /* subflavor */
1001         *vp++ = cpu_to_le32(ctx->gc_proc);              /* proc */
1002         *vp++ = cpu_to_le32(seqnum);                    /* seq */
1003         *vp++ = cpu_to_le32(PTLRPCS_GSS_SVC_INTEGRITY); /* service */
1004         vlen -= 5 * 4;
1005
1006         if (rawobj_serialize(&ctx->gc_wire_ctx, &vp, &vlen)) {
1007                 rc = -EIO;
1008                 goto out;
1009         }
1010         CDEBUG(D_SEC, "encoded wire_ctx length %d\n", ctx->gc_wire_ctx.len);
1011
1012         vpsave = vp++;  /* reserve for size */
1013         vlen -= 4;
1014
1015         mic.len = vlen;
1016         mic.data = (unsigned char *)vp;
1017
1018         CDEBUG(D_SEC, "reqbuf at %p, lmsg at %p, len %d, mic at %p, len %d\n",
1019                req->rq_reqbuf, lmsg.data, lmsg.len, mic.data, mic.len);
1020         major = kgss_get_mic(ctx->gc_gss_ctx, GSS_C_QOP_DEFAULT, &lmsg, &mic);
1021         if (major) {
1022                 CERROR("cred %p: req %p compute mic error, major %x\n",
1023                        cred, req, major);
1024                 rc = -EACCES;
1025                 goto out;
1026         }
1027
1028         *vpsave = cpu_to_le32(mic.len);
1029
1030         seclen = seclen - vlen + mic.len;
1031         buf_to_sec_hdr(req->rq_reqbuf)->sec_len = cpu_to_le32(seclen);
1032         req->rq_reqdata_len += size_round(seclen);
1033         CDEBUG(D_SEC, "msg size %d, checksum size %d, total sec size %d\n",
1034                lmsg.len, mic.len, seclen);
1035 out:
1036         gss_put_ctx(ctx);
1037         RETURN(rc);
1038 }
1039
1040 static int gss_cred_verify(struct ptlrpc_cred *cred,
1041                            struct ptlrpc_request *req)
1042 {
1043         struct gss_cred        *gcred;
1044         struct gss_cl_ctx      *ctx;
1045         struct ptlrpcs_wire_hdr *sec_hdr;
1046         rawobj_t                lmsg, mic;
1047         __u32                   *vp, vlen, subflavor, proc, seq, svc;
1048         __u32                   major, minor, rc;
1049         ENTRY;
1050
1051         LASSERT(req->rq_repbuf);
1052         LASSERT(req->rq_cred == cred);
1053
1054         sec_hdr = buf_to_sec_hdr(req->rq_repbuf);
1055         vp = (__u32 *) (req->rq_repbuf + sizeof(*sec_hdr) + sec_hdr->msg_len);
1056         vlen = sec_hdr->sec_len;
1057
1058         if (vlen < 7 * 4) {
1059                 CERROR("reply sec size %u too small\n", vlen);
1060                 RETURN(-EPROTO);
1061         }
1062
1063         if (*vp++ != cpu_to_le32(PTLRPC_SEC_GSS_VERSION)) {
1064                 CERROR("reply have different gss version\n");
1065                 RETURN(-EPROTO);
1066         }
1067         subflavor = le32_to_cpu(*vp++);
1068         proc = le32_to_cpu(*vp++);
1069         vlen -= 3 * 4;
1070
1071         switch (proc) {
1072         case PTLRPCS_GSS_PROC_DATA:
1073                 seq = le32_to_cpu(*vp++);
1074                 svc = le32_to_cpu(*vp++);
1075                 if (svc != PTLRPCS_GSS_SVC_INTEGRITY) {
1076                         CERROR("Unknown svc %d\n", svc);
1077                         RETURN(-EPROTO);
1078                 }
1079                 if (*vp++ != 0) {
1080                         CERROR("Unexpected ctx handle\n");
1081                         RETURN(-EPROTO);
1082                 }
1083                 mic.len = le32_to_cpu(*vp++);
1084                 vlen -= 4 * 4;
1085                 if (vlen < mic.len) {
1086                         CERROR("vlen %d, mic.len %d\n", vlen, mic.len);
1087                         RETURN(-EINVAL);
1088                 }
1089                 mic.data = (unsigned char *)vp;
1090
1091                 gcred = container_of(cred, struct gss_cred, gc_base);
1092                 ctx = gss_cred_get_ctx(cred);
1093                 LASSERT(ctx);
1094
1095                 lmsg.len = sec_hdr->msg_len;
1096                 lmsg.data = (__u8 *) buf_to_lustre_msg(req->rq_repbuf);
1097
1098                 major = kgss_verify_mic(ctx->gc_gss_ctx, &lmsg, &mic, NULL);
1099                 if (major != GSS_S_COMPLETE) {
1100                         CERROR("cred %p: req %p verify mic error: major %x\n",
1101                                cred, req, major);
1102
1103                         if (major == GSS_S_CREDENTIALS_EXPIRED ||
1104                             major == GSS_S_CONTEXT_EXPIRED) {
1105                                 ptlrpcs_cred_expire(cred);
1106                                 req->rq_ptlrpcs_restart = 1;
1107                                 rc = 0;
1108                         } else
1109                                 rc = -EINVAL;
1110
1111                         GOTO(proc_data_out, rc);
1112                 }
1113
1114                 req->rq_repmsg = (struct lustre_msg *) lmsg.data;
1115                 req->rq_replen = lmsg.len;
1116
1117                 /* here we could check the seq number is the same one
1118                  * we sent to server. but portals has prevent us from
1119                  * replay attack, so maybe we don't need check it again.
1120                  */
1121                 rc = 0;
1122 proc_data_out:
1123                 gss_put_ctx(ctx);
1124                 break;
1125         case PTLRPCS_GSS_PROC_ERR:
1126                 major = le32_to_cpu(*vp++);
1127                 minor = le32_to_cpu(*vp++);
1128                 /* server return NO_CONTEXT might be caused by context expire
1129                  * or server reboot/failover. we refresh the cred transparently
1130                  * to upper layer.
1131                  * In some cases, our gss handle is possible to be incidentally
1132                  * identical to another handle since the handle itself is not
1133                  * fully random. In krb5 case, the GSS_S_BAD_SIG will be
1134                  * returned, maybe other gss error for other mechanism. Here we
1135                  * only consider krb5 mech (FIXME) and try to establish new
1136                  * context.
1137                  */
1138                 if (major == GSS_S_NO_CONTEXT ||
1139                     major == GSS_S_BAD_SIG) {
1140                         CWARN("req %p: server report cred %p %s, expired?\n",
1141                                req, cred, (major == GSS_S_NO_CONTEXT) ?
1142                                            "NO_CONTEXT" : "BAD_SIG");
1143
1144                         ptlrpcs_cred_expire(cred);
1145                         req->rq_ptlrpcs_restart = 1;
1146                         rc = 0;
1147                 } else {
1148                         CERROR("req %p: unrecognized gss error (%x/%x)\n",
1149                                 req, major, minor);
1150                         rc = -EACCES;
1151                 }
1152                 break;
1153         default:
1154                 CERROR("unknown gss proc %d\n", proc);
1155                 rc = -EPROTO;
1156         }
1157
1158         RETURN(rc);
1159 }
1160
1161 static int gss_cred_seal(struct ptlrpc_cred *cred,
1162                          struct ptlrpc_request *req)
1163 {
1164         struct gss_cred         *gcred;
1165         struct gss_cl_ctx       *ctx;
1166         struct ptlrpcs_wire_hdr *sec_hdr;
1167         rawobj_buf_t             msg_buf;
1168         rawobj_t                 cipher_buf;
1169         __u32                   *vp, *vpsave, vlen, seclen;
1170         __u32                    major, seqnum, rc = 0;
1171         ENTRY;
1172
1173         LASSERT(req->rq_reqbuf);
1174         LASSERT(req->rq_cred == cred);
1175
1176         gcred = container_of(cred, struct gss_cred, gc_base);
1177         ctx = gss_cred_get_ctx(cred);
1178         if (!ctx) {
1179                 CERROR("cred %p("LPU64"/%u) invalidated?\n",
1180                         cred, cred->pc_pag, cred->pc_uid);
1181                 RETURN(-EPERM);
1182         }
1183
1184         vp = (__u32 *) (req->rq_reqbuf + sizeof(*sec_hdr));
1185         vlen = req->rq_reqbuf_len - sizeof(*sec_hdr);
1186         seclen = vlen;
1187
1188         if (vlen < 6 * 4 + size_round4(ctx->gc_wire_ctx.len)) {
1189                 CERROR("vlen %d, need %d\n",
1190                         vlen, 6 * 4 + size_round4(ctx->gc_wire_ctx.len));
1191                 rc = -EIO;
1192                 goto out;
1193         }
1194
1195         spin_lock(&ctx->gc_seq_lock);
1196         seqnum = ctx->gc_seq++;
1197         spin_unlock(&ctx->gc_seq_lock);
1198
1199         *vp++ = cpu_to_le32(PTLRPC_SEC_GSS_VERSION);    /* version */
1200         *vp++ = cpu_to_le32(PTLRPCS_FLVR_KRB5P);        /* subflavor */
1201         *vp++ = cpu_to_le32(ctx->gc_proc);              /* proc */
1202         *vp++ = cpu_to_le32(seqnum);                    /* seq */
1203         *vp++ = cpu_to_le32(PTLRPCS_GSS_SVC_PRIVACY);   /* service */
1204         vlen -= 5 * 4;
1205
1206         if (rawobj_serialize(&ctx->gc_wire_ctx, &vp, &vlen)) {
1207                 rc = -EIO;
1208                 goto out;
1209         }
1210         CDEBUG(D_SEC, "encoded wire_ctx length %d\n", ctx->gc_wire_ctx.len);
1211
1212         vpsave = vp++;  /* reserve for size */
1213         vlen -= 4;
1214
1215         msg_buf.buf = (__u8 *) req->rq_reqmsg - GSS_PRIVBUF_PREFIX_LEN;
1216         msg_buf.buflen = req->rq_reqlen + GSS_PRIVBUF_PREFIX_LEN +
1217                                           GSS_PRIVBUF_SUFFIX_LEN;
1218         msg_buf.dataoff = GSS_PRIVBUF_PREFIX_LEN;
1219         msg_buf.datalen = req->rq_reqlen;
1220
1221         cipher_buf.data = (__u8 *) vp;
1222         cipher_buf.len = vlen;
1223
1224         major = kgss_wrap(ctx->gc_gss_ctx, GSS_C_QOP_DEFAULT,
1225                           &msg_buf, &cipher_buf);
1226         if (major) {
1227                 CERROR("cred %p: error wrap: major 0x%x\n", cred, major);
1228                 GOTO(out, rc = -EINVAL);
1229         }
1230
1231         *vpsave = cpu_to_le32(cipher_buf.len);
1232
1233         seclen = seclen - vlen + cipher_buf.len;
1234         sec_hdr = buf_to_sec_hdr(req->rq_reqbuf);
1235         sec_hdr->sec_len = cpu_to_le32(seclen);
1236         req->rq_reqdata_len += size_round(seclen);
1237
1238         CDEBUG(D_SEC, "msg size %d, total sec size %d\n",
1239                req->rq_reqlen, seclen);
1240 out:
1241         gss_put_ctx(ctx);
1242         RETURN(rc);
1243 }
1244
1245 static int gss_cred_unseal(struct ptlrpc_cred *cred,
1246                            struct ptlrpc_request *req)
1247 {
1248         struct gss_cred        *gcred;
1249         struct gss_cl_ctx      *ctx;
1250         struct ptlrpcs_wire_hdr *sec_hdr;
1251         rawobj_t                cipher_text, plain_text;
1252         __u32                   *vp, vlen, subflavor, proc, seq, svc;
1253         __u32                   major, rc;
1254         ENTRY;
1255
1256         LASSERT(req->rq_repbuf);
1257         LASSERT(req->rq_cred == cred);
1258
1259         sec_hdr = buf_to_sec_hdr(req->rq_repbuf);
1260         if (sec_hdr->msg_len != 0) {
1261                 CERROR("unexpected msg_len %u\n", sec_hdr->msg_len);
1262                 RETURN(-EPROTO);
1263         }
1264
1265         vp = (__u32 *) (req->rq_repbuf + sizeof(*sec_hdr));
1266         vlen = sec_hdr->sec_len;
1267
1268         if (vlen < 7 * 4) {
1269                 CERROR("reply sec size %u too small\n", vlen);
1270                 RETURN(-EPROTO);
1271         }
1272
1273         if (*vp++ != cpu_to_le32(PTLRPC_SEC_GSS_VERSION)) {
1274                 CERROR("reply have different gss version\n");
1275                 RETURN(-EPROTO);
1276         }
1277         subflavor = le32_to_cpu(*vp++);
1278         proc = le32_to_cpu(*vp++);
1279         seq = le32_to_cpu(*vp++);
1280         svc = le32_to_cpu(*vp++);
1281         vlen -= 5 * 4;
1282
1283         switch (proc) {
1284         case PTLRPCS_GSS_PROC_DATA:
1285                 if (svc != PTLRPCS_GSS_SVC_PRIVACY) {
1286                         CERROR("Unknown svc %d\n", svc);
1287                         RETURN(-EPROTO);
1288                 }
1289                 if (*vp++ != 0) {
1290                         CERROR("Unexpected ctx handle\n");
1291                         RETURN(-EPROTO);
1292                 }
1293                 vlen -= 4;
1294
1295                 cipher_text.len = le32_to_cpu(*vp++);
1296                 cipher_text.data = (__u8 *) vp;
1297                 vlen -= 4;
1298
1299                 if (vlen < cipher_text.len) {
1300                         CERROR("cipher text to be %u while buf only %u\n",
1301                                 cipher_text.len, vlen);
1302                         RETURN(-EPROTO);
1303                 }
1304
1305                 plain_text = cipher_text;
1306
1307                 gcred = container_of(cred, struct gss_cred, gc_base);
1308                 ctx = gss_cred_get_ctx(cred);
1309                 LASSERT(ctx);
1310
1311                 major = kgss_unwrap(ctx->gc_gss_ctx, GSS_C_QOP_DEFAULT,
1312                                     &cipher_text, &plain_text);
1313                 if (major) {
1314                         CERROR("cred %p: error unwrap: major 0x%x\n",
1315                                cred, major);
1316
1317                         if (major == GSS_S_CREDENTIALS_EXPIRED ||
1318                             major == GSS_S_CONTEXT_EXPIRED) {
1319                                 ptlrpcs_cred_expire(cred);
1320                                 req->rq_ptlrpcs_restart = 1;
1321                                 rc = 0;
1322                         } else
1323                                 rc = -EINVAL;
1324
1325                         GOTO(proc_out, rc);
1326                 }
1327
1328                 req->rq_repmsg = (struct lustre_msg *) vp;
1329                 req->rq_replen = plain_text.len;
1330
1331                 rc = 0;
1332 proc_out:
1333                 gss_put_ctx(ctx);
1334                 break;
1335         default:
1336                 CERROR("unknown gss proc %d\n", proc);
1337                 rc = -EPROTO;
1338         }
1339
1340         RETURN(rc);
1341 }
1342
1343 static void destroy_gss_context(struct ptlrpc_cred *cred)
1344 {
1345         struct ptlrpcs_wire_hdr *hdr;
1346         struct lustre_msg       *lmsg;
1347         struct gss_cred         *gcred;
1348         struct ptlrpc_request    req;
1349         struct obd_import       *imp;
1350         __u32                   *vp, lmsg_size;
1351         struct ptlrpc_request   *raw_req = NULL;
1352         const int                repbuf_len = 256;
1353         char                    *repbuf;
1354         int                      replen, rc;
1355         ENTRY;
1356
1357         imp = cred->pc_sec->ps_import;
1358         LASSERT(imp);
1359
1360         if (test_bit(PTLRPC_CRED_ERROR_BIT, &cred->pc_flags) ||
1361             !test_bit(PTLRPC_CRED_UPTODATE_BIT, &cred->pc_flags)) {
1362                 CDEBUG(D_SEC, "Destroy dead cred %p(%u@%s)\n",
1363                        cred, cred->pc_uid, imp->imp_target_uuid.uuid);
1364                 EXIT;
1365                 return;
1366         }
1367
1368         might_sleep();
1369
1370         /* cred's refcount is 0, steal one */
1371         atomic_inc(&cred->pc_refcount);
1372
1373         gcred = container_of(cred, struct gss_cred, gc_base);
1374         gcred->gc_ctx->gc_proc = PTLRPCS_GSS_PROC_DESTROY;
1375
1376         CDEBUG(D_SEC, "client destroy gss cred %p(%u@%s)\n",
1377                gcred, cred->pc_uid, imp->imp_target_uuid.uuid);
1378
1379         lmsg_size = lustre_msg_size(0, NULL);
1380         req.rq_req_secflvr = cred->pc_sec->ps_flavor;
1381         req.rq_cred = cred;
1382         req.rq_reqbuf_len = sizeof(*hdr) + lmsg_size +
1383                             ptlrpcs_est_req_payload(&req, lmsg_size);
1384
1385         OBD_ALLOC(req.rq_reqbuf, req.rq_reqbuf_len);
1386         if (!req.rq_reqbuf) {
1387                 CERROR("Fail to alloc reqbuf, cancel anyway\n");
1388                 atomic_dec(&cred->pc_refcount);
1389                 EXIT;
1390                 return;
1391         }
1392
1393         /* wire hdr */
1394         hdr = buf_to_sec_hdr(req.rq_reqbuf);
1395         hdr->flavor  = cpu_to_le32(PTLRPCS_FLVR_GSS_AUTH);
1396         hdr->msg_len = cpu_to_le32(lmsg_size);
1397         hdr->sec_len = cpu_to_le32(0);
1398
1399         /* lustre message */
1400         lmsg = buf_to_lustre_msg(req.rq_reqbuf);
1401         lustre_init_msg(lmsg, 0, NULL, NULL);
1402         lmsg->handle   = imp->imp_remote_handle;
1403         lmsg->type     = PTL_RPC_MSG_REQUEST;
1404         lmsg->opc      = SEC_FINI;
1405         lmsg->flags    = 0;
1406         lmsg->conn_cnt = imp->imp_conn_cnt;
1407         /* add this for randomize */
1408         get_random_bytes(&lmsg->last_xid, sizeof(lmsg->last_xid));
1409         get_random_bytes(&lmsg->transno, sizeof(lmsg->transno));
1410
1411         vp = (__u32 *) req.rq_reqbuf;
1412
1413         req.rq_cred = cred;
1414         req.rq_reqmsg = buf_to_lustre_msg(req.rq_reqbuf);
1415         req.rq_reqlen = lmsg_size;
1416         req.rq_reqdata_len = sizeof(*hdr) + lmsg_size;
1417
1418         if (gss_cred_sign(cred, &req)) {
1419                 CERROR("failed to sign, cancel anyway\n");
1420                 atomic_dec(&cred->pc_refcount);
1421                 goto exit;
1422         }
1423         atomic_dec(&cred->pc_refcount);
1424
1425         OBD_ALLOC(repbuf, repbuf_len);
1426         if (!repbuf)
1427                 goto exit;
1428
1429         raw_req = ptl_do_rawrpc(imp, req.rq_reqbuf, req.rq_reqbuf_len,
1430                                 req.rq_reqdata_len, repbuf, repbuf_len, &replen,
1431                                 SECFINI_RPC_TIMEOUT, &rc);
1432         if (!raw_req)
1433                 OBD_FREE(repbuf, repbuf_len);
1434
1435 exit:
1436         if (raw_req == NULL)
1437                 OBD_FREE(req.rq_reqbuf, req.rq_reqbuf_len);
1438         else
1439                 rawrpc_req_finished(raw_req);
1440         EXIT;
1441 }
1442
1443 static void gss_cred_destroy(struct ptlrpc_cred *cred)
1444 {
1445         struct gss_cred *gcred;
1446         ENTRY;
1447
1448         LASSERT(cred);
1449         LASSERT(!atomic_read(&cred->pc_refcount));
1450
1451         gcred = container_of(cred, struct gss_cred, gc_base);
1452         if (gcred->gc_ctx) {
1453                 destroy_gss_context(cred);
1454                 gss_put_ctx(gcred->gc_ctx);
1455         }
1456
1457         CDEBUG(D_SEC, "sec.gss %p: destroy cred %p\n", cred->pc_sec, gcred);
1458
1459         OBD_FREE(gcred, sizeof(*gcred));
1460         EXIT;
1461 }
1462
1463 static struct ptlrpc_credops gss_credops = {
1464         .refresh        = gss_cred_refresh,
1465         .match          = gss_cred_match,
1466         .sign           = gss_cred_sign,
1467         .verify         = gss_cred_verify,
1468         .seal           = gss_cred_seal,
1469         .unseal         = gss_cred_unseal,
1470         .destroy        = gss_cred_destroy,
1471 };
1472
1473 #ifdef __KERNEL__
1474 /*******************************************
1475  * rpc_pipe APIs                           *
1476  *******************************************/
1477 static ssize_t
1478 gss_pipe_upcall(struct file *filp, struct rpc_pipe_msg *msg,
1479                 char *dst, size_t buflen)
1480 {
1481         char *data = (char *)msg->data + msg->copied;
1482         ssize_t mlen = msg->len;
1483         ssize_t left;
1484         ENTRY;
1485
1486         if (mlen > buflen)
1487                 mlen = buflen;
1488         left = copy_to_user(dst, data, mlen);
1489         if (left < 0) {
1490                 msg->errno = left;
1491                 RETURN(left);
1492         }
1493         mlen -= left;
1494         msg->copied += mlen;
1495         msg->errno = 0;
1496         RETURN(mlen);
1497 }
1498
1499 static ssize_t
1500 gss_pipe_downcall(struct file *filp, const char *src, size_t mlen)
1501 {
1502         char *buf;
1503         const int bufsize = 1024;
1504         rawobj_t obj;
1505         struct inode *inode = filp->f_dentry->d_inode;
1506         struct rpc_inode *rpci = RPC_I(inode);
1507         struct obd_import *import;
1508         struct ptlrpc_sec *sec;
1509         struct gss_sec *gsec;
1510         char *obdname;
1511         struct gss_api_mech *mech;
1512         struct vfs_cred vcred = { 0 };
1513         struct ptlrpc_cred *cred;
1514         struct gss_upcall_msg *gss_msg;
1515         struct gss_upcall_msg_data gmd = { 0 };
1516         struct gss_cl_ctx *ctx = NULL;
1517         ssize_t left;
1518         int err, gss_err;
1519         ENTRY;
1520
1521         if (mlen > bufsize) {
1522                 CERROR("mlen %ld > bufsize %d\n", (long)mlen, bufsize);
1523                 RETURN(-ENOSPC);
1524         }
1525
1526         OBD_ALLOC(buf, bufsize);
1527         if (!buf) {
1528                 CERROR("alloc mem failed\n");
1529                 RETURN(-ENOMEM);
1530         }
1531
1532         left = copy_from_user(buf, src, mlen);
1533         if (left)
1534                 GOTO(err_free, err = -EFAULT);
1535
1536         obj.data = (unsigned char *)buf;
1537         obj.len = mlen;
1538
1539         LASSERT(rpci->private);
1540         gsec = (struct gss_sec *)rpci->private;
1541         sec = &gsec->gs_base;
1542         LASSERT(sec->ps_import);
1543         import = class_import_get(sec->ps_import);
1544         LASSERT(import->imp_obd);
1545         obdname = import->imp_obd->obd_name;
1546         mech = gsec->gs_mech;
1547
1548         err = gss_parse_init_downcall(mech, &obj, &ctx, &gmd, &gss_err);
1549         if (err)
1550                 CERROR("parse init downcall err %d\n", err);
1551
1552         vcred.vc_pag = gmd.gum_pag;
1553         vcred.vc_uid = gmd.gum_uid;
1554
1555         cred = ptlrpcs_cred_lookup(sec, &vcred);
1556         if (!cred) {
1557                 CWARN("didn't find cred for uid %u\n", vcred.vc_uid);
1558                 GOTO(err, err = -EINVAL);
1559         }
1560
1561         if (err || gss_err) {
1562                 set_bit(PTLRPC_CRED_DEAD_BIT, &cred->pc_flags);
1563                 if (err != -ERESTART || gss_err != 0)
1564                         set_bit(PTLRPC_CRED_ERROR_BIT, &cred->pc_flags);
1565                 CERROR("cred %p: rpc err %d, gss err 0x%x, fatal %d\n",
1566                        cred, err, gss_err,
1567                        (test_bit(PTLRPC_CRED_ERROR_BIT, &cred->pc_flags) != 0));
1568         } else {
1569                 CDEBUG(D_SEC, "get initial ctx:\n");
1570                 gss_cred_set_ctx(cred, ctx);
1571         }
1572
1573         spin_lock(&gsec->gs_lock);
1574         gss_msg = gss_find_upcall(gsec, obdname, &gmd);
1575         if (gss_msg) {
1576                 gss_unhash_msg_nolock(gss_msg);
1577                 spin_unlock(&gsec->gs_lock);
1578                 gss_release_msg(gss_msg);
1579         } else
1580                 spin_unlock(&gsec->gs_lock);
1581
1582         ptlrpcs_cred_put(cred, 1);
1583         class_import_put(import);
1584         OBD_FREE(buf, bufsize);
1585         RETURN(mlen);
1586 err:
1587         if (ctx)
1588                 gss_destroy_ctx(ctx);
1589         class_import_put(import);
1590 err_free:
1591         OBD_FREE(buf, bufsize);
1592         CDEBUG(D_SEC, "gss_pipe_downcall returning %d\n", err);
1593         RETURN(err);
1594 }
1595
1596 static
1597 void gss_pipe_destroy_msg(struct rpc_pipe_msg *msg)
1598 {
1599         struct gss_upcall_msg *gmsg;
1600         static unsigned long ratelimit;
1601         ENTRY;
1602
1603         LASSERT(list_empty(&msg->list));
1604
1605         if (msg->errno >= 0) {
1606                 EXIT;
1607                 return;
1608         }
1609
1610         gmsg = container_of(msg, struct gss_upcall_msg, gum_base);
1611         CDEBUG(D_SEC, "destroy gmsg %p\n", gmsg);
1612         LASSERT(atomic_read(&gmsg->gum_refcount) > 0);
1613         atomic_inc(&gmsg->gum_refcount);
1614         gss_unhash_msg(gmsg);
1615         if (msg->errno == -ETIMEDOUT || msg->errno == -EPIPE) {
1616                 unsigned long now = get_seconds();
1617                 if (time_after(now, ratelimit)) {
1618                         CWARN("sec.gss upcall timed out.\n"
1619                               "Please check user daemon is running!\n");
1620                         ratelimit = now + 15;
1621                 }
1622         }
1623         gss_release_msg(gmsg);
1624         EXIT;
1625 }
1626
1627 static
1628 void gss_pipe_release(struct inode *inode)
1629 {
1630         struct rpc_inode *rpci = RPC_I(inode);
1631         struct ptlrpc_sec *sec;
1632         struct gss_sec *gsec;
1633         ENTRY;
1634
1635         gsec = (struct gss_sec *)rpci->private;
1636         sec = &gsec->gs_base;
1637         spin_lock(&gsec->gs_lock);
1638         while (!list_empty(&gsec->gs_upcalls)) {
1639                 struct gss_upcall_msg *gmsg;
1640
1641                 gmsg = list_entry(gsec->gs_upcalls.next,
1642                                   struct gss_upcall_msg, gum_list);
1643                 LASSERT(list_empty(&gmsg->gum_base.list));
1644                 gmsg->gum_base.errno = -EPIPE;
1645                 atomic_inc(&gmsg->gum_refcount);
1646                 gss_unhash_msg_nolock(gmsg);
1647                 gss_release_msg(gmsg);
1648         }
1649         spin_unlock(&gsec->gs_lock);
1650         EXIT;
1651 }
1652
1653 static struct rpc_pipe_ops gss_upcall_ops = {
1654         .upcall         = gss_pipe_upcall,
1655         .downcall       = gss_pipe_downcall,
1656         .destroy_msg    = gss_pipe_destroy_msg,
1657         .release_pipe   = gss_pipe_release,
1658 };
1659 #endif /* __KERNEL__ */
1660
1661 /*********************************************
1662  * GSS security APIs                         *
1663  *********************************************/
1664
1665 static
1666 struct ptlrpc_sec* gss_create_sec(__u32 flavor,
1667                                   const char *pipe_dir,
1668                                   void *pipe_data)
1669 {
1670         struct gss_sec *gsec;
1671         struct ptlrpc_sec *sec;
1672         uid_t save_uid;
1673
1674 #ifdef __KERNEL__
1675         char *pos;
1676         int   pipepath_len;
1677 #endif
1678         ENTRY;
1679
1680         LASSERT(SEC_FLAVOR_MAJOR(flavor) == PTLRPCS_FLVR_MAJOR_GSS);
1681
1682         OBD_ALLOC(gsec, sizeof(*gsec));
1683         if (!gsec) {
1684                 CERROR("can't alloc gsec\n");
1685                 RETURN(NULL);
1686         }
1687
1688         gsec->gs_mech = kgss_subflavor_to_mech(SEC_FLAVOR_SUB(flavor));
1689         if (!gsec->gs_mech) {
1690                 CERROR("subflavor 0x%x not found\n", flavor);
1691                 goto err_free;
1692         }
1693
1694         /* initialize gss sec */
1695 #ifdef __KERNEL__
1696         INIT_LIST_HEAD(&gsec->gs_upcalls);
1697         spin_lock_init(&gsec->gs_lock);
1698
1699         pipepath_len = strlen(LUSTRE_PIPEDIR) + strlen(pipe_dir) +
1700                        strlen(gsec->gs_mech->gm_name) + 3;
1701         OBD_ALLOC(gsec->gs_pipepath, pipepath_len);
1702         if (!gsec->gs_pipepath)
1703                 goto err_mech_put;
1704
1705         /* pipe rpc require root permission */
1706         save_uid = current->fsuid;
1707         current->fsuid = 0;
1708
1709         sprintf(gsec->gs_pipepath, LUSTRE_PIPEDIR"/%s", pipe_dir);
1710         if (IS_ERR(rpc_mkdir(gsec->gs_pipepath, NULL))) {
1711                 CERROR("can't make pipedir %s\n", gsec->gs_pipepath);
1712                 goto err_free_path;
1713         }
1714
1715         sprintf(gsec->gs_pipepath, LUSTRE_PIPEDIR"/%s/%s", pipe_dir,
1716                 gsec->gs_mech->gm_name);
1717         gsec->gs_depipe = rpc_mkpipe(gsec->gs_pipepath, gsec,
1718                                      &gss_upcall_ops, RPC_PIPE_WAIT_FOR_OPEN);
1719         if (IS_ERR(gsec->gs_depipe)) {
1720                 CERROR("failed to make rpc_pipe %s: %ld\n",
1721                         gsec->gs_pipepath, PTR_ERR(gsec->gs_depipe));
1722                 goto err_rmdir;
1723         }
1724         CDEBUG(D_SEC, "gss sec %p, pipe path %s\n", gsec, gsec->gs_pipepath);
1725 #endif
1726
1727         sec = &gsec->gs_base;
1728         sec->ps_expire = GSS_CREDCACHE_EXPIRE;
1729         sec->ps_nextgc = get_seconds() + sec->ps_expire;
1730         sec->ps_flags = 0;
1731
1732         current->fsuid = save_uid;
1733
1734         CDEBUG(D_SEC, "Create sec.gss %p\n", gsec);
1735         RETURN(sec);
1736
1737 #ifdef __KERNEL__
1738 err_rmdir:
1739         pos = strrchr(gsec->gs_pipepath, '/');
1740         LASSERT(pos);
1741         *pos = 0;
1742         rpc_rmdir(gsec->gs_pipepath);
1743 err_free_path:
1744         current->fsuid = save_uid;
1745         OBD_FREE(gsec->gs_pipepath, pipepath_len);
1746 err_mech_put:
1747 #endif
1748         kgss_mech_put(gsec->gs_mech);
1749 err_free:
1750         OBD_FREE(gsec, sizeof(*gsec));
1751         RETURN(NULL);
1752 }
1753
1754 static
1755 void gss_destroy_sec(struct ptlrpc_sec *sec)
1756 {
1757         struct gss_sec *gsec;
1758 #ifdef __KERNEL__
1759         char *pos;
1760         int   pipepath_len;
1761 #endif
1762         ENTRY;
1763
1764         gsec = container_of(sec, struct gss_sec, gs_base);
1765         CDEBUG(D_SEC, "Destroy sec.gss %p\n", gsec);
1766
1767         LASSERT(gsec->gs_mech);
1768         LASSERT(!atomic_read(&sec->ps_refcount));
1769         LASSERT(!atomic_read(&sec->ps_credcount));
1770 #ifdef __KERNEL__
1771         pipepath_len = strlen(gsec->gs_pipepath) + 1;
1772         rpc_unlink(gsec->gs_pipepath);
1773         pos = strrchr(gsec->gs_pipepath, '/');
1774         LASSERT(pos);
1775         *pos = 0;
1776         rpc_rmdir(gsec->gs_pipepath);
1777         OBD_FREE(gsec->gs_pipepath, pipepath_len);
1778 #endif
1779
1780         kgss_mech_put(gsec->gs_mech);
1781         OBD_FREE(gsec, sizeof(*gsec));
1782         EXIT;
1783 }
1784
1785 static
1786 struct ptlrpc_cred * gss_create_cred(struct ptlrpc_sec *sec,
1787                                      struct vfs_cred *vcred)
1788 {
1789         struct gss_cred *gcred;
1790         struct ptlrpc_cred *cred;
1791         ENTRY;
1792
1793         OBD_ALLOC(gcred, sizeof(*gcred));
1794         if (!gcred)
1795                 RETURN(NULL);
1796
1797         cred = &gcred->gc_base;
1798         INIT_LIST_HEAD(&cred->pc_hash);
1799         atomic_set(&cred->pc_refcount, 0);
1800         cred->pc_sec = sec;
1801         cred->pc_ops = &gss_credops;
1802         cred->pc_expire = 0;
1803         cred->pc_flags = 0;
1804         cred->pc_pag = vcred->vc_pag;
1805         cred->pc_uid = vcred->vc_uid;
1806         CDEBUG(D_SEC, "create a gss cred at %p("LPU64"/%u)\n",
1807                cred, vcred->vc_pag, vcred->vc_uid);
1808
1809         RETURN(cred);
1810 }
1811
1812 static int gss_estimate_payload(struct ptlrpc_sec *sec,
1813                                 struct ptlrpc_request *req,
1814                                 int msgsize)
1815 {
1816         switch (SEC_FLAVOR_SVC(req->rq_req_secflvr)) {
1817         case PTLRPCS_SVC_AUTH:
1818                 return GSS_MAX_AUTH_PAYLOAD;
1819         case PTLRPCS_SVC_PRIV:
1820                 return size_round16(GSS_MAX_AUTH_PAYLOAD + msgsize +
1821                                     GSS_PRIVBUF_PREFIX_LEN +
1822                                     GSS_PRIVBUF_SUFFIX_LEN);
1823         default:
1824                 LBUG();
1825                 return 0;
1826         }
1827 }
1828
1829 static int gss_alloc_reqbuf(struct ptlrpc_sec *sec,
1830                             struct ptlrpc_request *req,
1831                             int lmsg_size)
1832 {
1833         int msg_payload, sec_payload;
1834         int privacy, rc;
1835         ENTRY;
1836
1837         /* In PRIVACY mode, lustre message is always 0 (already encoded into
1838          * security payload).
1839          */
1840         privacy = (SEC_FLAVOR_SVC(req->rq_req_secflvr) == PTLRPCS_SVC_PRIV);
1841         msg_payload = privacy ? 0 : lmsg_size;
1842         sec_payload = gss_estimate_payload(sec, req, lmsg_size);
1843
1844         rc = sec_alloc_reqbuf(sec, req, msg_payload, sec_payload);
1845         if (rc)
1846                 return rc;
1847
1848         if (privacy) {
1849                 int buflen = lmsg_size + GSS_PRIVBUF_PREFIX_LEN +
1850                              GSS_PRIVBUF_SUFFIX_LEN;
1851                 char *buf;
1852
1853                 OBD_ALLOC(buf, buflen);
1854                 if (!buf) {
1855                         CERROR("Fail to alloc %d\n", buflen);
1856                         sec_free_reqbuf(sec, req);
1857                         RETURN(-ENOMEM);
1858                 }
1859                 req->rq_reqmsg = (struct lustre_msg *)
1860                                         (buf + GSS_PRIVBUF_PREFIX_LEN);
1861         }
1862
1863         RETURN(0);
1864 }
1865
1866 static void gss_free_reqbuf(struct ptlrpc_sec *sec,
1867                             struct ptlrpc_request *req)
1868 {
1869         char *buf;
1870         int privacy;
1871         ENTRY;
1872
1873         LASSERT(req->rq_reqmsg);
1874         LASSERT(req->rq_reqlen);
1875
1876         privacy = SEC_FLAVOR_SVC(req->rq_req_secflvr) == PTLRPCS_SVC_PRIV;
1877         if (privacy) {
1878                 buf = (char *) req->rq_reqmsg - GSS_PRIVBUF_PREFIX_LEN;
1879                 LASSERT(buf < req->rq_reqbuf ||
1880                         buf >= req->rq_reqbuf + req->rq_reqbuf_len);
1881                 OBD_FREE(buf, req->rq_reqlen + GSS_PRIVBUF_PREFIX_LEN +
1882                               GSS_PRIVBUF_SUFFIX_LEN);
1883                 req->rq_reqmsg = NULL;
1884         }
1885
1886         sec_free_reqbuf(sec, req);
1887 }
1888
1889 static struct ptlrpc_secops gss_secops = {
1890         .create_sec             = gss_create_sec,
1891         .destroy_sec            = gss_destroy_sec,
1892         .create_cred            = gss_create_cred,
1893         .est_req_payload        = gss_estimate_payload,
1894         .est_rep_payload        = gss_estimate_payload,
1895         .alloc_reqbuf           = gss_alloc_reqbuf,
1896         .free_reqbuf            = gss_free_reqbuf,
1897 };
1898
1899 static struct ptlrpc_sec_type gss_type = {
1900         .pst_owner      = THIS_MODULE,
1901         .pst_name       = "sec.gss",
1902         .pst_inst       = ATOMIC_INIT(0),
1903         .pst_flavor     = PTLRPCS_FLVR_MAJOR_GSS,
1904         .pst_ops        = &gss_secops,
1905 };
1906
1907 extern int
1908 (*lustre_secinit_downcall_handler)(char *buffer, unsigned long count);
1909
1910 int __init ptlrpcs_gss_init(void)
1911 {
1912         int rc;
1913
1914         rc = ptlrpcs_register(&gss_type);
1915         if (rc)
1916                 return rc;
1917
1918 #ifdef __KERNEL__
1919         gss_svc_init();
1920
1921         rc = PTR_ERR(rpc_mkdir(LUSTRE_PIPEDIR, NULL));
1922         if (IS_ERR((void *)rc) && rc != -EEXIST) {
1923                 CERROR("fail to make rpcpipedir for lustre\n");
1924                 gss_svc_exit();
1925                 ptlrpcs_unregister(&gss_type);
1926                 return -1;
1927         }
1928         rc = 0;
1929 #else
1930 #endif
1931         rc = init_kerberos_module();
1932         if (rc) {
1933                 ptlrpcs_unregister(&gss_type);
1934         }
1935
1936         lustre_secinit_downcall_handler = gss_send_secinit_rpc;
1937
1938         return rc;
1939 }
1940
1941 #ifdef __KERNEL__
1942 static void __exit ptlrpcs_gss_exit(void)
1943 {
1944         lustre_secinit_downcall_handler = NULL;
1945
1946         cleanup_kerberos_module();
1947         rpc_rmdir(LUSTRE_PIPEDIR);
1948         gss_svc_exit();
1949         ptlrpcs_unregister(&gss_type);
1950 }
1951 #endif
1952
1953 MODULE_AUTHOR("Cluster File Systems, Inc. <info@clusterfs.com>");
1954 MODULE_DESCRIPTION("GSS Security module for Lustre");
1955 MODULE_LICENSE("GPL");
1956
1957 module_init(ptlrpcs_gss_init);
1958 module_exit(ptlrpcs_gss_exit);