Whamcloud - gitweb
74b0d33a608f0dac678a616d33a54b4bfc61f736
[fs/lustre-release.git] / lustre / ptlrpc / 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  *
6  * Copyright (c) 2007, 2010, Oracle and/or its affiliates. All rights reserved.
7  *
8  * Author: Eric Mei <ericm@clusterfs.com>
9  */
10
11 /*
12  * linux/net/sunrpc/auth_gss.c
13  *
14  * RPCSEC_GSS client authentication.
15  *
16  *  Copyright (c) 2000 The Regents of the University of Michigan.
17  *  All rights reserved.
18  *
19  *  Dug Song       <dugsong@monkey.org>
20  *  Andy Adamson   <andros@umich.edu>
21  *
22  *  Redistribution and use in source and binary forms, with or without
23  *  modification, are permitted provided that the following conditions
24  *  are met:
25  *
26  *  1. Redistributions of source code must retain the above copyright
27  *     notice, this list of conditions and the following disclaimer.
28  *  2. Redistributions in binary form must reproduce the above copyright
29  *     notice, this list of conditions and the following disclaimer in the
30  *     documentation and/or other materials provided with the distribution.
31  *  3. Neither the name of the University nor the names of its
32  *     contributors may be used to endorse or promote products derived
33  *     from this software without specific prior written permission.
34  *
35  *  THIS SOFTWARE IS PROVIDED ``AS IS'' AND ANY EXPRESS OR IMPLIED
36  *  WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES OF
37  *  MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
38  *  DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
39  *  FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
40  *  CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
41  *  SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR
42  *  BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF
43  *  LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING
44  *  NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
45  *  SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
46  *
47  */
48
49 #ifndef EXPORT_SYMTAB
50 # define EXPORT_SYMTAB
51 #endif
52 #define DEBUG_SUBSYSTEM S_SEC
53 #ifdef __KERNEL__
54 #include <linux/init.h>
55 #include <linux/module.h>
56 #include <linux/slab.h>
57 #include <linux/dcache.h>
58 #include <linux/fs.h>
59 #include <linux/mutex.h>
60 #include <asm/atomic.h>
61 #else
62 #include <liblustre.h>
63 #endif
64
65 #include <obd.h>
66 #include <obd_class.h>
67 #include <obd_support.h>
68 #include <obd_cksum.h>
69 #include <lustre/lustre_idl.h>
70 #include <lustre_net.h>
71 #include <lustre_import.h>
72 #include <lustre_sec.h>
73
74 #include "gss_err.h"
75 #include "gss_internal.h"
76 #include "gss_api.h"
77
78 #include <linux/crypto.h>
79
80 /*
81  * early reply have fixed size, respectively in privacy and integrity mode.
82  * so we calculate them only once.
83  */
84 static int gss_at_reply_off_integ;
85 static int gss_at_reply_off_priv;
86
87
88 static inline int msg_last_segidx(struct lustre_msg *msg)
89 {
90         LASSERT(msg->lm_bufcount > 0);
91         return msg->lm_bufcount - 1;
92 }
93 static inline int msg_last_seglen(struct lustre_msg *msg)
94 {
95         return msg->lm_buflens[msg_last_segidx(msg)];
96 }
97
98 /********************************************
99  * wire data swabber                        *
100  ********************************************/
101
102 static
103 void gss_header_swabber(struct gss_header *ghdr)
104 {
105         __swab32s(&ghdr->gh_flags);
106         __swab32s(&ghdr->gh_proc);
107         __swab32s(&ghdr->gh_seq);
108         __swab32s(&ghdr->gh_svc);
109         __swab32s(&ghdr->gh_pad1);
110         __swab32s(&ghdr->gh_handle.len);
111 }
112
113 struct gss_header *gss_swab_header(struct lustre_msg *msg, int segment,
114                                    int swabbed)
115 {
116         struct gss_header *ghdr;
117
118         ghdr = lustre_msg_buf(msg, segment, sizeof(*ghdr));
119         if (ghdr == NULL)
120                 return NULL;
121
122         if (swabbed)
123                 gss_header_swabber(ghdr);
124
125         if (sizeof(*ghdr) + ghdr->gh_handle.len > msg->lm_buflens[segment]) {
126                 CERROR("gss header has length %d, now %u received\n",
127                        (int) sizeof(*ghdr) + ghdr->gh_handle.len,
128                        msg->lm_buflens[segment]);
129                 return NULL;
130         }
131
132         return ghdr;
133 }
134
135 #if 0
136 static
137 void gss_netobj_swabber(netobj_t *obj)
138 {
139         __swab32s(&obj->len);
140 }
141
142 netobj_t *gss_swab_netobj(struct lustre_msg *msg, int segment)
143 {
144         netobj_t  *obj;
145
146         obj = lustre_swab_buf(msg, segment, sizeof(*obj), gss_netobj_swabber);
147         if (obj && sizeof(*obj) + obj->len > msg->lm_buflens[segment]) {
148                 CERROR("netobj require length %u but only %u received\n",
149                        (unsigned int) sizeof(*obj) + obj->len,
150                        msg->lm_buflens[segment]);
151                 return NULL;
152         }
153
154         return obj;
155 }
156 #endif
157
158 /*
159  * payload should be obtained from mechanism. but currently since we
160  * only support kerberos, we could simply use fixed value.
161  * krb5 "meta" data:
162  *  - krb5 header:      16
163  *  - krb5 checksum:    20
164  *
165  * for privacy mode, payload also include the cipher text which has the same
166  * size as plain text, plus possible confounder, padding both at maximum cipher
167  * block size.
168  */
169 #define GSS_KRB5_INTEG_MAX_PAYLOAD      (40)
170
171 static inline
172 int gss_mech_payload(struct gss_ctx *mechctx, int msgsize, int privacy)
173 {
174         if (privacy)
175                 return GSS_KRB5_INTEG_MAX_PAYLOAD + 16 + 16 + 16 + msgsize;
176         else
177                 return GSS_KRB5_INTEG_MAX_PAYLOAD;
178 }
179
180 /*
181  * return signature size, otherwise < 0 to indicate error
182  */
183 static int gss_sign_msg(struct lustre_msg *msg,
184                         struct gss_ctx *mechctx,
185                         enum lustre_sec_part sp,
186                         __u32 flags, __u32 proc, __u32 seq, __u32 svc,
187                         rawobj_t *handle)
188 {
189         struct gss_header      *ghdr;
190         rawobj_t                text[4], mic;
191         int                     textcnt, max_textcnt, mic_idx;
192         __u32                   major;
193
194         LASSERT(msg->lm_bufcount >= 2);
195
196         /* gss hdr */
197         LASSERT(msg->lm_buflens[0] >=
198                 sizeof(*ghdr) + (handle ? handle->len : 0));
199         ghdr = lustre_msg_buf(msg, 0, 0);
200
201         ghdr->gh_version = PTLRPC_GSS_VERSION;
202         ghdr->gh_sp = (__u8) sp;
203         ghdr->gh_flags = flags;
204         ghdr->gh_proc = proc;
205         ghdr->gh_seq = seq;
206         ghdr->gh_svc = svc;
207         if (!handle) {
208                 /* fill in a fake one */
209                 ghdr->gh_handle.len = 0;
210         } else {
211                 ghdr->gh_handle.len = handle->len;
212                 memcpy(ghdr->gh_handle.data, handle->data, handle->len);
213         }
214
215         /* no actual signature for null mode */
216         if (svc == SPTLRPC_SVC_NULL)
217                 return lustre_msg_size_v2(msg->lm_bufcount, msg->lm_buflens);
218
219         /* MIC */
220         mic_idx = msg_last_segidx(msg);
221         max_textcnt = (svc == SPTLRPC_SVC_AUTH) ? 1 : mic_idx;
222
223         for (textcnt = 0; textcnt < max_textcnt; textcnt++) {
224                 text[textcnt].len = msg->lm_buflens[textcnt];
225                 text[textcnt].data = lustre_msg_buf(msg, textcnt, 0);
226         }
227
228         mic.len = msg->lm_buflens[mic_idx];
229         mic.data = lustre_msg_buf(msg, mic_idx, 0);
230
231         major = lgss_get_mic(mechctx, textcnt, text, 0, NULL, &mic);
232         if (major != GSS_S_COMPLETE) {
233                 CERROR("fail to generate MIC: %08x\n", major);
234                 return -EPERM;
235         }
236         LASSERT(mic.len <= msg->lm_buflens[mic_idx]);
237
238         return lustre_shrink_msg(msg, mic_idx, mic.len, 0);
239 }
240
241 /*
242  * return gss error
243  */
244 static
245 __u32 gss_verify_msg(struct lustre_msg *msg,
246                      struct gss_ctx *mechctx,
247                      __u32 svc)
248 {
249         rawobj_t        text[4], mic;
250         int             textcnt, max_textcnt;
251         int             mic_idx;
252         __u32           major;
253
254         LASSERT(msg->lm_bufcount >= 2);
255
256         if (svc == SPTLRPC_SVC_NULL)
257                 return GSS_S_COMPLETE;
258
259         mic_idx = msg_last_segidx(msg);
260         max_textcnt = (svc == SPTLRPC_SVC_AUTH) ? 1 : mic_idx;
261
262         for (textcnt = 0; textcnt < max_textcnt; textcnt++) {
263                 text[textcnt].len = msg->lm_buflens[textcnt];
264                 text[textcnt].data = lustre_msg_buf(msg, textcnt, 0);
265         }
266
267         mic.len = msg->lm_buflens[mic_idx];
268         mic.data = lustre_msg_buf(msg, mic_idx, 0);
269
270         major = lgss_verify_mic(mechctx, textcnt, text, 0, NULL, &mic);
271         if (major != GSS_S_COMPLETE)
272                 CERROR("mic verify error: %08x\n", major);
273
274         return major;
275 }
276
277 /*
278  * return gss error code
279  */
280 static
281 __u32 gss_unseal_msg(struct gss_ctx *mechctx,
282                    struct lustre_msg *msgbuf,
283                    int *msg_len, int msgbuf_len)
284 {
285         rawobj_t                 clear_obj, hdrobj, token;
286         __u8                    *clear_buf;
287         int                      clear_buflen;
288         __u32                    major;
289         ENTRY;
290
291         if (msgbuf->lm_bufcount != 2) {
292                 CERROR("invalid bufcount %d\n", msgbuf->lm_bufcount);
293                 RETURN(GSS_S_FAILURE);
294         }
295
296         /* allocate a temporary clear text buffer, same sized as token,
297          * we assume the final clear text size <= token size */
298         clear_buflen = lustre_msg_buflen(msgbuf, 1);
299         OBD_ALLOC(clear_buf, clear_buflen);
300         if (!clear_buf)
301                 RETURN(GSS_S_FAILURE);
302
303         /* buffer objects */
304         hdrobj.len = lustre_msg_buflen(msgbuf, 0);
305         hdrobj.data = lustre_msg_buf(msgbuf, 0, 0);
306         token.len = lustre_msg_buflen(msgbuf, 1);
307         token.data = lustre_msg_buf(msgbuf, 1, 0);
308         clear_obj.len = clear_buflen;
309         clear_obj.data = clear_buf;
310
311         major = lgss_unwrap(mechctx, &hdrobj, &token, &clear_obj);
312         if (major != GSS_S_COMPLETE) {
313                 CERROR("unwrap message error: %08x\n", major);
314                 GOTO(out_free, major = GSS_S_FAILURE);
315         }
316         LASSERT(clear_obj.len <= clear_buflen);
317         LASSERT(clear_obj.len <= msgbuf_len);
318
319         /* now the decrypted message */
320         memcpy(msgbuf, clear_obj.data, clear_obj.len);
321         *msg_len = clear_obj.len;
322
323         major = GSS_S_COMPLETE;
324 out_free:
325         OBD_FREE(clear_buf, clear_buflen);
326         RETURN(major);
327 }
328
329 /********************************************
330  * gss client context manipulation helpers  *
331  ********************************************/
332
333 int cli_ctx_expire(struct ptlrpc_cli_ctx *ctx)
334 {
335         LASSERT(cfs_atomic_read(&ctx->cc_refcount));
336
337         if (!cfs_test_and_set_bit(PTLRPC_CTX_DEAD_BIT, &ctx->cc_flags)) {
338                 if (!ctx->cc_early_expire)
339                         cfs_clear_bit(PTLRPC_CTX_UPTODATE_BIT, &ctx->cc_flags);
340
341                 CWARN("ctx %p(%u->%s) get expired: %lu(%+lds)\n",
342                       ctx, ctx->cc_vcred.vc_uid, sec2target_str(ctx->cc_sec),
343                       ctx->cc_expire,
344                       ctx->cc_expire == 0 ? 0 :
345                       cfs_time_sub(ctx->cc_expire, cfs_time_current_sec()));
346
347                 sptlrpc_cli_ctx_wakeup(ctx);
348                 return 1;
349         }
350
351         return 0;
352 }
353
354 /*
355  * return 1 if the context is dead.
356  */
357 int cli_ctx_check_death(struct ptlrpc_cli_ctx *ctx)
358 {
359         if (unlikely(cli_ctx_is_dead(ctx)))
360                 return 1;
361
362         /* expire is 0 means never expire. a newly created gss context
363          * which during upcall may has 0 expiration */
364         if (ctx->cc_expire == 0)
365                 return 0;
366
367         /* check real expiration */
368         if (cfs_time_after(ctx->cc_expire, cfs_time_current_sec()))
369                 return 0;
370
371         cli_ctx_expire(ctx);
372         return 1;
373 }
374
375 void gss_cli_ctx_uptodate(struct gss_cli_ctx *gctx)
376 {
377         struct ptlrpc_cli_ctx  *ctx = &gctx->gc_base;
378         unsigned long           ctx_expiry;
379
380         if (lgss_inquire_context(gctx->gc_mechctx, &ctx_expiry)) {
381                 CERROR("ctx %p(%u): unable to inquire, expire it now\n",
382                        gctx, ctx->cc_vcred.vc_uid);
383                 ctx_expiry = 1; /* make it expired now */
384         }
385
386         ctx->cc_expire = gss_round_ctx_expiry(ctx_expiry,
387                                               ctx->cc_sec->ps_flvr.sf_flags);
388
389         /* At this point this ctx might have been marked as dead by
390          * someone else, in which case nobody will make further use
391          * of it. we don't care, and mark it UPTODATE will help
392          * destroying server side context when it be destroied. */
393         cfs_set_bit(PTLRPC_CTX_UPTODATE_BIT, &ctx->cc_flags);
394
395         if (sec_is_reverse(ctx->cc_sec)) {
396                 CWARN("server installed reverse ctx %p idx "LPX64", "
397                       "expiry %lu(%+lds)\n", ctx,
398                       gss_handle_to_u64(&gctx->gc_handle),
399                       ctx->cc_expire, ctx->cc_expire - cfs_time_current_sec());
400         } else {
401                 CWARN("client refreshed ctx %p idx "LPX64" (%u->%s), "
402                       "expiry %lu(%+lds)\n", ctx,
403                       gss_handle_to_u64(&gctx->gc_handle),
404                       ctx->cc_vcred.vc_uid, sec2target_str(ctx->cc_sec),
405                       ctx->cc_expire, ctx->cc_expire - cfs_time_current_sec());
406
407                 /* install reverse svc ctx for root context */
408                 if (ctx->cc_vcred.vc_uid == 0)
409                         gss_sec_install_rctx(ctx->cc_sec->ps_import,
410                                              ctx->cc_sec, ctx);
411         }
412
413         sptlrpc_cli_ctx_wakeup(ctx);
414 }
415
416 static void gss_cli_ctx_finalize(struct gss_cli_ctx *gctx)
417 {
418         LASSERT(gctx->gc_base.cc_sec);
419
420         if (gctx->gc_mechctx) {
421                 lgss_delete_sec_context(&gctx->gc_mechctx);
422                 gctx->gc_mechctx = NULL;
423         }
424
425         if (!rawobj_empty(&gctx->gc_svc_handle)) {
426                 /* forward ctx: mark buddy reverse svcctx soon-expire. */
427                 if (!sec_is_reverse(gctx->gc_base.cc_sec) &&
428                     !rawobj_empty(&gctx->gc_svc_handle))
429                         gss_svc_upcall_expire_rvs_ctx(&gctx->gc_svc_handle);
430
431                 rawobj_free(&gctx->gc_svc_handle);
432         }
433
434         rawobj_free(&gctx->gc_handle);
435 }
436
437 /*
438  * Based on sequence number algorithm as specified in RFC 2203.
439  *
440  * modified for our own problem: arriving request has valid sequence number,
441  * but unwrapping request might cost a long time, after that its sequence
442  * are not valid anymore (fall behind the window). It rarely happen, mostly
443  * under extreme load.
444  *
445  * note we should not check sequence before verify the integrity of incoming
446  * request, because just one attacking request with high sequence number might
447  * cause all following request be dropped.
448  *
449  * so here we use a multi-phase approach: prepare 2 sequence windows,
450  * "main window" for normal sequence and "back window" for fall behind sequence.
451  * and 3-phase checking mechanism:
452  *  0 - before integrity verification, perform a initial sequence checking in
453  *      main window, which only try and don't actually set any bits. if the
454  *      sequence is high above the window or fit in the window and the bit
455  *      is 0, then accept and proceed to integrity verification. otherwise
456  *      reject this sequence.
457  *  1 - after integrity verification, check in main window again. if this
458  *      sequence is high above the window or fit in the window and the bit
459  *      is 0, then set the bit and accept; if it fit in the window but bit
460  *      already set, then reject; if it fall behind the window, then proceed
461  *      to phase 2.
462  *  2 - check in back window. if it is high above the window or fit in the
463  *      window and the bit is 0, then set the bit and accept. otherwise reject.
464  *
465  * return value:
466  *   1: looks like a replay
467  *   0: is ok
468  *  -1: is a replay
469  *
470  * note phase 0 is necessary, because otherwise replay attacking request of
471  * sequence which between the 2 windows can't be detected.
472  *
473  * this mechanism can't totally solve the problem, but could help much less
474  * number of valid requests be dropped.
475  */
476 static
477 int gss_do_check_seq(unsigned long *window, __u32 win_size, __u32 *max_seq,
478                      __u32 seq_num, int phase)
479 {
480         LASSERT(phase >= 0 && phase <= 2);
481
482         if (seq_num > *max_seq) {
483                 /*
484                  * 1. high above the window
485                  */
486                 if (phase == 0)
487                         return 0;
488
489                 if (seq_num >= *max_seq + win_size) {
490                         memset(window, 0, win_size / 8);
491                         *max_seq = seq_num;
492                 } else {
493                         while(*max_seq < seq_num) {
494                                 (*max_seq)++;
495                                 __clear_bit((*max_seq) % win_size, window);
496                         }
497                 }
498                 __set_bit(seq_num % win_size, window);
499         } else if (seq_num + win_size <= *max_seq) {
500                 /*
501                  * 2. low behind the window
502                  */
503                 if (phase == 0 || phase == 2)
504                         goto replay;
505
506                 CWARN("seq %u is %u behind (size %d), check backup window\n",
507                       seq_num, *max_seq - win_size - seq_num, win_size);
508                 return 1;
509         } else {
510                 /*
511                  * 3. fit into the window
512                  */
513                 switch (phase) {
514                 case 0:
515                         if (cfs_test_bit(seq_num % win_size, window))
516                                 goto replay;
517                         break;
518                 case 1:
519                 case 2:
520                      if (__test_and_set_bit(seq_num % win_size, window))
521                                 goto replay;
522                         break;
523                 }
524         }
525
526         return 0;
527
528 replay:
529         CERROR("seq %u (%s %s window) is a replay: max %u, winsize %d\n",
530                seq_num,
531                seq_num + win_size > *max_seq ? "in" : "behind",
532                phase == 2 ? "backup " : "main",
533                *max_seq, win_size);
534         return -1;
535 }
536
537 /*
538  * Based on sequence number algorithm as specified in RFC 2203.
539  *
540  * if @set == 0: initial check, don't set any bit in window
541  * if @sec == 1: final check, set bit in window
542  */
543 int gss_check_seq_num(struct gss_svc_seq_data *ssd, __u32 seq_num, int set)
544 {
545         int rc = 0;
546
547         cfs_spin_lock(&ssd->ssd_lock);
548
549         if (set == 0) {
550                 /*
551                  * phase 0 testing
552                  */
553                 rc = gss_do_check_seq(ssd->ssd_win_main, GSS_SEQ_WIN_MAIN,
554                                       &ssd->ssd_max_main, seq_num, 0);
555                 if (unlikely(rc))
556                         gss_stat_oos_record_svc(0, 1);
557         } else {
558                 /*
559                  * phase 1 checking main window
560                  */
561                 rc = gss_do_check_seq(ssd->ssd_win_main, GSS_SEQ_WIN_MAIN,
562                                       &ssd->ssd_max_main, seq_num, 1);
563                 switch (rc) {
564                 case -1:
565                         gss_stat_oos_record_svc(1, 1);
566                         /* fall through */
567                 case 0:
568                         goto exit;
569                 }
570                 /*
571                  * phase 2 checking back window
572                  */
573                 rc = gss_do_check_seq(ssd->ssd_win_back, GSS_SEQ_WIN_BACK,
574                                       &ssd->ssd_max_back, seq_num, 2);
575                 if (rc)
576                         gss_stat_oos_record_svc(2, 1);
577                 else
578                         gss_stat_oos_record_svc(2, 0);
579         }
580 exit:
581         cfs_spin_unlock(&ssd->ssd_lock);
582         return rc;
583 }
584
585 /***************************************
586  * cred APIs                           *
587  ***************************************/
588
589 static inline int gss_cli_payload(struct ptlrpc_cli_ctx *ctx,
590                                   int msgsize, int privacy)
591 {
592         return gss_mech_payload(NULL, msgsize, privacy);
593 }
594
595 static int gss_cli_bulk_payload(struct ptlrpc_cli_ctx *ctx,
596                                 struct sptlrpc_flavor *flvr,
597                                 int reply, int read)
598 {
599         int     payload = sizeof(struct ptlrpc_bulk_sec_desc);
600
601         LASSERT(SPTLRPC_FLVR_BULK_TYPE(flvr->sf_rpc) == SPTLRPC_BULK_DEFAULT);
602
603         if ((!reply && !read) || (reply && read)) {
604                 switch (SPTLRPC_FLVR_BULK_SVC(flvr->sf_rpc)) {
605                 case SPTLRPC_BULK_SVC_NULL:
606                         break;
607                 case SPTLRPC_BULK_SVC_INTG:
608                         payload += gss_cli_payload(ctx, 0, 0);
609                         break;
610                 case SPTLRPC_BULK_SVC_PRIV:
611                         payload += gss_cli_payload(ctx, 0, 1);
612                         break;
613                 case SPTLRPC_BULK_SVC_AUTH:
614                 default:
615                         LBUG();
616                 }
617         }
618
619         return payload;
620 }
621
622 int gss_cli_ctx_match(struct ptlrpc_cli_ctx *ctx, struct vfs_cred *vcred)
623 {
624         return (ctx->cc_vcred.vc_uid == vcred->vc_uid);
625 }
626
627 void gss_cli_ctx_flags2str(unsigned long flags, char *buf, int bufsize)
628 {
629         buf[0] = '\0';
630
631         if (flags & PTLRPC_CTX_NEW)
632                 strncat(buf, "new,", bufsize);
633         if (flags & PTLRPC_CTX_UPTODATE)
634                 strncat(buf, "uptodate,", bufsize);
635         if (flags & PTLRPC_CTX_DEAD)
636                 strncat(buf, "dead,", bufsize);
637         if (flags & PTLRPC_CTX_ERROR)
638                 strncat(buf, "error,", bufsize);
639         if (flags & PTLRPC_CTX_CACHED)
640                 strncat(buf, "cached,", bufsize);
641         if (flags & PTLRPC_CTX_ETERNAL)
642                 strncat(buf, "eternal,", bufsize);
643         if (buf[0] == '\0')
644                 strncat(buf, "-,", bufsize);
645
646         buf[strlen(buf) - 1] = '\0';
647 }
648
649 int gss_cli_ctx_sign(struct ptlrpc_cli_ctx *ctx,
650                      struct ptlrpc_request *req)
651 {
652         struct gss_cli_ctx      *gctx = ctx2gctx(ctx);
653         __u32                    flags = 0, seq, svc;
654         int                      rc;
655         ENTRY;
656
657         LASSERT(req->rq_reqbuf);
658         LASSERT(req->rq_reqbuf->lm_bufcount >= 2);
659         LASSERT(req->rq_cli_ctx == ctx);
660
661         /* nothing to do for context negotiation RPCs */
662         if (req->rq_ctx_init)
663                 RETURN(0);
664
665         svc = SPTLRPC_FLVR_SVC(req->rq_flvr.sf_rpc);
666         if (req->rq_pack_bulk)
667                 flags |= LUSTRE_GSS_PACK_BULK;
668         if (req->rq_pack_udesc)
669                 flags |= LUSTRE_GSS_PACK_USER;
670
671 redo:
672         seq = cfs_atomic_inc_return(&gctx->gc_seq);
673
674         rc = gss_sign_msg(req->rq_reqbuf, gctx->gc_mechctx,
675                           ctx->cc_sec->ps_part,
676                           flags, gctx->gc_proc, seq, svc,
677                           &gctx->gc_handle);
678         if (rc < 0)
679                 RETURN(rc);
680
681         /* gss_sign_msg() msg might take long time to finish, in which period
682          * more rpcs could be wrapped up and sent out. if we found too many
683          * of them we should repack this rpc, because sent it too late might
684          * lead to the sequence number fall behind the window on server and
685          * be dropped. also applies to gss_cli_ctx_seal().
686          *
687          * Note: null mode dosen't check sequence number. */
688         if (svc != SPTLRPC_SVC_NULL &&
689             cfs_atomic_read(&gctx->gc_seq) - seq > GSS_SEQ_REPACK_THRESHOLD) {
690                 int behind = cfs_atomic_read(&gctx->gc_seq) - seq;
691
692                 gss_stat_oos_record_cli(behind);
693                 CWARN("req %p: %u behind, retry signing\n", req, behind);
694                 goto redo;
695         }
696
697         req->rq_reqdata_len = rc;
698         RETURN(0);
699 }
700
701 static
702 int gss_cli_ctx_handle_err_notify(struct ptlrpc_cli_ctx *ctx,
703                                   struct ptlrpc_request *req,
704                                   struct gss_header *ghdr)
705 {
706         struct gss_err_header *errhdr;
707         int rc;
708
709         LASSERT(ghdr->gh_proc == PTLRPC_GSS_PROC_ERR);
710
711         errhdr = (struct gss_err_header *) ghdr;
712
713         CWARN("req x"LPU64"/t"LPU64", ctx %p idx "LPX64"(%u->%s): "
714               "%sserver respond (%08x/%08x)\n",
715               req->rq_xid, req->rq_transno, ctx,
716               gss_handle_to_u64(&ctx2gctx(ctx)->gc_handle),
717               ctx->cc_vcred.vc_uid, sec2target_str(ctx->cc_sec),
718               sec_is_reverse(ctx->cc_sec) ? "reverse" : "",
719               errhdr->gh_major, errhdr->gh_minor);
720
721         /* context fini rpc, let it failed */
722         if (req->rq_ctx_fini) {
723                 CWARN("context fini rpc failed\n");
724                 return -EINVAL;
725         }
726
727         /* reverse sec, just return error, don't expire this ctx because it's
728          * crucial to callback rpcs. note if the callback rpc failed because
729          * of bit flip during network transfer, the client will be evicted
730          * directly. so more gracefully we probably want let it retry for
731          * number of times. */
732         if (sec_is_reverse(ctx->cc_sec))
733                 return -EINVAL;
734
735         if (errhdr->gh_major != GSS_S_NO_CONTEXT &&
736             errhdr->gh_major != GSS_S_BAD_SIG)
737                 return -EACCES;
738
739         /* server return NO_CONTEXT might be caused by context expire
740          * or server reboot/failover. we try to refresh a new ctx which
741          * be transparent to upper layer.
742          *
743          * In some cases, our gss handle is possible to be incidentally
744          * identical to another handle since the handle itself is not
745          * fully random. In krb5 case, the GSS_S_BAD_SIG will be
746          * returned, maybe other gss error for other mechanism.
747          *
748          * if we add new mechanism, make sure the correct error are
749          * returned in this case. */
750         CWARN("%s: server might lost the context, retrying\n",
751               errhdr->gh_major == GSS_S_NO_CONTEXT ?  "NO_CONTEXT" : "BAD_SIG");
752
753         sptlrpc_cli_ctx_expire(ctx);
754
755         /* we need replace the ctx right here, otherwise during
756          * resent we'll hit the logic in sptlrpc_req_refresh_ctx()
757          * which keep the ctx with RESEND flag, thus we'll never
758          * get rid of this ctx. */
759         rc = sptlrpc_req_replace_dead_ctx(req);
760         if (rc == 0)
761                 req->rq_resend = 1;
762
763         return rc;
764 }
765
766 int gss_cli_ctx_verify(struct ptlrpc_cli_ctx *ctx,
767                        struct ptlrpc_request *req)
768 {
769         struct gss_cli_ctx     *gctx;
770         struct gss_header      *ghdr, *reqhdr;
771         struct lustre_msg      *msg = req->rq_repdata;
772         __u32                   major;
773         int                     pack_bulk, swabbed, rc = 0;
774         ENTRY;
775
776         LASSERT(req->rq_cli_ctx == ctx);
777         LASSERT(msg);
778
779         gctx = container_of(ctx, struct gss_cli_ctx, gc_base);
780
781         /* special case for context negotiation, rq_repmsg/rq_replen actually
782          * are not used currently. but early reply always be treated normally */
783         if (req->rq_ctx_init && !req->rq_early) {
784                 req->rq_repmsg = lustre_msg_buf(msg, 1, 0);
785                 req->rq_replen = msg->lm_buflens[1];
786                 RETURN(0);
787         }
788
789         if (msg->lm_bufcount < 2 || msg->lm_bufcount > 4) {
790                 CERROR("unexpected bufcount %u\n", msg->lm_bufcount);
791                 RETURN(-EPROTO);
792         }
793
794         swabbed = ptlrpc_rep_need_swab(req);
795
796         ghdr = gss_swab_header(msg, 0, swabbed);
797         if (ghdr == NULL) {
798                 CERROR("can't decode gss header\n");
799                 RETURN(-EPROTO);
800         }
801
802         /* sanity checks */
803         reqhdr = lustre_msg_buf(msg, 0, sizeof(*reqhdr));
804         LASSERT(reqhdr);
805
806         if (ghdr->gh_version != reqhdr->gh_version) {
807                 CERROR("gss version %u mismatch, expect %u\n",
808                        ghdr->gh_version, reqhdr->gh_version);
809                 RETURN(-EPROTO);
810         }
811
812         switch (ghdr->gh_proc) {
813         case PTLRPC_GSS_PROC_DATA:
814                 pack_bulk = ghdr->gh_flags & LUSTRE_GSS_PACK_BULK;
815
816                 if (!req->rq_early && !equi(req->rq_pack_bulk == 1, pack_bulk)){
817                         CERROR("%s bulk flag in reply\n",
818                                req->rq_pack_bulk ? "missing" : "unexpected");
819                         RETURN(-EPROTO);
820                 }
821
822                 if (ghdr->gh_seq != reqhdr->gh_seq) {
823                         CERROR("seqnum %u mismatch, expect %u\n",
824                                ghdr->gh_seq, reqhdr->gh_seq);
825                         RETURN(-EPROTO);
826                 }
827
828                 if (ghdr->gh_svc != reqhdr->gh_svc) {
829                         CERROR("svc %u mismatch, expect %u\n",
830                                ghdr->gh_svc, reqhdr->gh_svc);
831                         RETURN(-EPROTO);
832                 }
833
834                 if (swabbed)
835                         gss_header_swabber(ghdr);
836
837                 major = gss_verify_msg(msg, gctx->gc_mechctx, reqhdr->gh_svc);
838                 if (major != GSS_S_COMPLETE) {
839                         CERROR("failed to verify reply: %x\n", major);
840                         RETURN(-EPERM);
841                 }
842
843                 if (req->rq_early && reqhdr->gh_svc == SPTLRPC_SVC_NULL) {
844                         __u32 cksum;
845
846                         cksum = crc32_le(!(__u32) 0,
847                                          lustre_msg_buf(msg, 1, 0),
848                                          lustre_msg_buflen(msg, 1));
849                         if (cksum != msg->lm_cksum) {
850                                 CWARN("early reply checksum mismatch: "
851                                       "%08x != %08x\n", cksum, msg->lm_cksum);
852                                 RETURN(-EPROTO);
853                         }
854                 }
855
856                 if (pack_bulk) {
857                         /* bulk checksum is right after the lustre msg */
858                         if (msg->lm_bufcount < 3) {
859                                 CERROR("Invalid reply bufcount %u\n",
860                                        msg->lm_bufcount);
861                                 RETURN(-EPROTO);
862                         }
863
864                         rc = bulk_sec_desc_unpack(msg, 2, swabbed);
865                         if (rc) {
866                                 CERROR("unpack bulk desc: %d\n", rc);
867                                 RETURN(rc);
868                         }
869                 }
870
871                 req->rq_repmsg = lustre_msg_buf(msg, 1, 0);
872                 req->rq_replen = msg->lm_buflens[1];
873                 break;
874         case PTLRPC_GSS_PROC_ERR:
875                 if (req->rq_early) {
876                         CERROR("server return error with early reply\n");
877                         rc = -EPROTO;
878                 } else {
879                         rc = gss_cli_ctx_handle_err_notify(ctx, req, ghdr);
880                 }
881                 break;
882         default:
883                 CERROR("unknown gss proc %d\n", ghdr->gh_proc);
884                 rc = -EPROTO;
885         }
886
887         RETURN(rc);
888 }
889
890 int gss_cli_ctx_seal(struct ptlrpc_cli_ctx *ctx,
891                      struct ptlrpc_request *req)
892 {
893         struct gss_cli_ctx      *gctx;
894         rawobj_t                 hdrobj, msgobj, token;
895         struct gss_header       *ghdr;
896         __u32                    buflens[2], major;
897         int                      wiresize, rc;
898         ENTRY;
899
900         LASSERT(req->rq_clrbuf);
901         LASSERT(req->rq_cli_ctx == ctx);
902         LASSERT(req->rq_reqlen);
903
904         gctx = container_of(ctx, struct gss_cli_ctx, gc_base);
905
906         /* final clear data length */
907         req->rq_clrdata_len = lustre_msg_size_v2(req->rq_clrbuf->lm_bufcount,
908                                                  req->rq_clrbuf->lm_buflens);
909
910         /* calculate wire data length */
911         buflens[0] = PTLRPC_GSS_HEADER_SIZE;
912         buflens[1] = gss_cli_payload(&gctx->gc_base, req->rq_clrdata_len, 1);
913         wiresize = lustre_msg_size_v2(2, buflens);
914
915         /* allocate wire buffer */
916         if (req->rq_pool) {
917                 /* pre-allocated */
918                 LASSERT(req->rq_reqbuf);
919                 LASSERT(req->rq_reqbuf != req->rq_clrbuf);
920                 LASSERT(req->rq_reqbuf_len >= wiresize);
921         } else {
922                 OBD_ALLOC(req->rq_reqbuf, wiresize);
923                 if (!req->rq_reqbuf)
924                         RETURN(-ENOMEM);
925                 req->rq_reqbuf_len = wiresize;
926         }
927
928         lustre_init_msg_v2(req->rq_reqbuf, 2, buflens, NULL);
929         req->rq_reqbuf->lm_secflvr = req->rq_flvr.sf_rpc;
930
931         /* gss header */
932         ghdr = lustre_msg_buf(req->rq_reqbuf, 0, 0);
933         ghdr->gh_version = PTLRPC_GSS_VERSION;
934         ghdr->gh_sp = (__u8) ctx->cc_sec->ps_part;
935         ghdr->gh_flags = 0;
936         ghdr->gh_proc = gctx->gc_proc;
937         ghdr->gh_svc = SPTLRPC_SVC_PRIV;
938         ghdr->gh_handle.len = gctx->gc_handle.len;
939         memcpy(ghdr->gh_handle.data, gctx->gc_handle.data, gctx->gc_handle.len);
940         if (req->rq_pack_bulk)
941                 ghdr->gh_flags |= LUSTRE_GSS_PACK_BULK;
942         if (req->rq_pack_udesc)
943                 ghdr->gh_flags |= LUSTRE_GSS_PACK_USER;
944
945 redo:
946         ghdr->gh_seq = cfs_atomic_inc_return(&gctx->gc_seq);
947
948         /* buffer objects */
949         hdrobj.len = PTLRPC_GSS_HEADER_SIZE;
950         hdrobj.data = (__u8 *) ghdr;
951         msgobj.len = req->rq_clrdata_len;
952         msgobj.data = (__u8 *) req->rq_clrbuf;
953         token.len = lustre_msg_buflen(req->rq_reqbuf, 1);
954         token.data = lustre_msg_buf(req->rq_reqbuf, 1, 0);
955
956         major = lgss_wrap(gctx->gc_mechctx, &hdrobj, &msgobj,
957                           req->rq_clrbuf_len, &token);
958         if (major != GSS_S_COMPLETE) {
959                 CERROR("priv: wrap message error: %08x\n", major);
960                 GOTO(err_free, rc = -EPERM);
961         }
962         LASSERT(token.len <= buflens[1]);
963
964         /* see explain in gss_cli_ctx_sign() */
965         if (unlikely(cfs_atomic_read(&gctx->gc_seq) - ghdr->gh_seq >
966                      GSS_SEQ_REPACK_THRESHOLD)) {
967                 int behind = cfs_atomic_read(&gctx->gc_seq) - ghdr->gh_seq;
968
969                 gss_stat_oos_record_cli(behind);
970                 CWARN("req %p: %u behind, retry sealing\n", req, behind);
971
972                 ghdr->gh_seq = cfs_atomic_inc_return(&gctx->gc_seq);
973                 goto redo;
974         }
975
976         /* now set the final wire data length */
977         req->rq_reqdata_len = lustre_shrink_msg(req->rq_reqbuf, 1, token.len,0);
978         RETURN(0);
979
980 err_free:
981         if (!req->rq_pool) {
982                 OBD_FREE(req->rq_reqbuf, req->rq_reqbuf_len);
983                 req->rq_reqbuf = NULL;
984                 req->rq_reqbuf_len = 0;
985         }
986         RETURN(rc);
987 }
988
989 int gss_cli_ctx_unseal(struct ptlrpc_cli_ctx *ctx,
990                        struct ptlrpc_request *req)
991 {
992         struct gss_cli_ctx      *gctx;
993         struct gss_header       *ghdr;
994         struct lustre_msg       *msg = req->rq_repdata;
995         int                      msglen, pack_bulk, swabbed, rc;
996         __u32                    major;
997         ENTRY;
998
999         LASSERT(req->rq_cli_ctx == ctx);
1000         LASSERT(req->rq_ctx_init == 0);
1001         LASSERT(msg);
1002
1003         gctx = container_of(ctx, struct gss_cli_ctx, gc_base);
1004         swabbed = ptlrpc_rep_need_swab(req);
1005
1006         ghdr = gss_swab_header(msg, 0, swabbed);
1007         if (ghdr == NULL) {
1008                 CERROR("can't decode gss header\n");
1009                 RETURN(-EPROTO);
1010         }
1011
1012         /* sanity checks */
1013         if (ghdr->gh_version != PTLRPC_GSS_VERSION) {
1014                 CERROR("gss version %u mismatch, expect %u\n",
1015                        ghdr->gh_version, PTLRPC_GSS_VERSION);
1016                 RETURN(-EPROTO);
1017         }
1018
1019         switch (ghdr->gh_proc) {
1020         case PTLRPC_GSS_PROC_DATA:
1021                 pack_bulk = ghdr->gh_flags & LUSTRE_GSS_PACK_BULK;
1022
1023                 if (!req->rq_early && !equi(req->rq_pack_bulk == 1, pack_bulk)){
1024                         CERROR("%s bulk flag in reply\n",
1025                                req->rq_pack_bulk ? "missing" : "unexpected");
1026                         RETURN(-EPROTO);
1027                 }
1028
1029                 if (swabbed)
1030                         gss_header_swabber(ghdr);
1031
1032                 /* use rq_repdata_len as buffer size, which assume unseal
1033                  * doesn't need extra memory space. for precise control, we'd
1034                  * better calculate out actual buffer size as
1035                  * (repbuf_len - offset - repdata_len) */
1036                 major = gss_unseal_msg(gctx->gc_mechctx, msg,
1037                                        &msglen, req->rq_repdata_len);
1038                 if (major != GSS_S_COMPLETE) {
1039                         CERROR("failed to unwrap reply: %x\n", major);
1040                         rc = -EPERM;
1041                         break;
1042                 }
1043
1044                 swabbed = __lustre_unpack_msg(msg, msglen);
1045                 if (swabbed < 0) {
1046                         CERROR("Failed to unpack after decryption\n");
1047                         RETURN(-EPROTO);
1048                 }
1049
1050                 if (msg->lm_bufcount < 1) {
1051                         CERROR("Invalid reply buffer: empty\n");
1052                         RETURN(-EPROTO);
1053                 }
1054
1055                 if (pack_bulk) {
1056                         if (msg->lm_bufcount < 2) {
1057                                 CERROR("bufcount %u: missing bulk sec desc\n",
1058                                        msg->lm_bufcount);
1059                                 RETURN(-EPROTO);
1060                         }
1061
1062                         /* bulk checksum is the last segment */
1063                         if (bulk_sec_desc_unpack(msg, msg->lm_bufcount - 1,
1064                                                  swabbed))
1065                                 RETURN(-EPROTO);
1066                 }
1067
1068                 req->rq_repmsg = lustre_msg_buf(msg, 0, 0);
1069                 req->rq_replen = msg->lm_buflens[0];
1070
1071                 rc = 0;
1072                 break;
1073         case PTLRPC_GSS_PROC_ERR:
1074                 if (req->rq_early) {
1075                         CERROR("server return error with early reply\n");
1076                         rc = -EPROTO;
1077                 } else {
1078                         rc = gss_cli_ctx_handle_err_notify(ctx, req, ghdr);
1079                 }
1080                 break;
1081         default:
1082                 CERROR("unexpected proc %d\n", ghdr->gh_proc);
1083                 rc = -EPERM;
1084         }
1085
1086         RETURN(rc);
1087 }
1088
1089 /*********************************************
1090  * reverse context installation              *
1091  *********************************************/
1092
1093 static inline
1094 int gss_install_rvs_svc_ctx(struct obd_import *imp,
1095                             struct gss_sec *gsec,
1096                             struct gss_cli_ctx *gctx)
1097 {
1098         return gss_svc_upcall_install_rvs_ctx(imp, gsec, gctx);
1099 }
1100
1101 /*********************************************
1102  * GSS security APIs                         *
1103  *********************************************/
1104 int gss_sec_create_common(struct gss_sec *gsec,
1105                           struct ptlrpc_sec_policy *policy,
1106                           struct obd_import *imp,
1107                           struct ptlrpc_svc_ctx *svcctx,
1108                           struct sptlrpc_flavor *sf)
1109 {
1110         struct ptlrpc_sec   *sec;
1111
1112         LASSERT(imp);
1113         LASSERT(SPTLRPC_FLVR_POLICY(sf->sf_rpc) == SPTLRPC_POLICY_GSS);
1114
1115         gsec->gs_mech = lgss_subflavor_to_mech(
1116                                 SPTLRPC_FLVR_BASE_SUB(sf->sf_rpc));
1117         if (!gsec->gs_mech) {
1118                 CERROR("gss backend 0x%x not found\n",
1119                        SPTLRPC_FLVR_BASE_SUB(sf->sf_rpc));
1120                 return -EOPNOTSUPP;
1121         }
1122
1123         cfs_spin_lock_init(&gsec->gs_lock);
1124         gsec->gs_rvs_hdl = 0ULL;
1125
1126         /* initialize upper ptlrpc_sec */
1127         sec = &gsec->gs_base;
1128         sec->ps_policy = policy;
1129         cfs_atomic_set(&sec->ps_refcount, 0);
1130         cfs_atomic_set(&sec->ps_nctx, 0);
1131         sec->ps_id = sptlrpc_get_next_secid();
1132         sec->ps_flvr = *sf;
1133         sec->ps_import = class_import_get(imp);
1134         cfs_spin_lock_init(&sec->ps_lock);
1135         CFS_INIT_LIST_HEAD(&sec->ps_gc_list);
1136
1137         if (!svcctx) {
1138                 sec->ps_gc_interval = GSS_GC_INTERVAL;
1139         } else {
1140                 LASSERT(sec_is_reverse(sec));
1141
1142                 /* never do gc on reverse sec */
1143                 sec->ps_gc_interval = 0;
1144         }
1145
1146         if (SPTLRPC_FLVR_BULK_SVC(sec->ps_flvr.sf_rpc) == SPTLRPC_BULK_SVC_PRIV)
1147                 sptlrpc_enc_pool_add_user();
1148
1149         CDEBUG(D_SEC, "create %s%s@%p\n", (svcctx ? "reverse " : ""),
1150                policy->sp_name, gsec);
1151         return 0;
1152 }
1153
1154 void gss_sec_destroy_common(struct gss_sec *gsec)
1155 {
1156         struct ptlrpc_sec      *sec = &gsec->gs_base;
1157         ENTRY;
1158
1159         LASSERT(sec->ps_import);
1160         LASSERT(cfs_atomic_read(&sec->ps_refcount) == 0);
1161         LASSERT(cfs_atomic_read(&sec->ps_nctx) == 0);
1162
1163         if (gsec->gs_mech) {
1164                 lgss_mech_put(gsec->gs_mech);
1165                 gsec->gs_mech = NULL;
1166         }
1167
1168         class_import_put(sec->ps_import);
1169
1170         if (SPTLRPC_FLVR_BULK_SVC(sec->ps_flvr.sf_rpc) == SPTLRPC_BULK_SVC_PRIV)
1171                 sptlrpc_enc_pool_del_user();
1172
1173         EXIT;
1174 }
1175
1176 void gss_sec_kill(struct ptlrpc_sec *sec)
1177 {
1178         sec->ps_dying = 1;
1179 }
1180
1181 int gss_cli_ctx_init_common(struct ptlrpc_sec *sec,
1182                             struct ptlrpc_cli_ctx *ctx,
1183                             struct ptlrpc_ctx_ops *ctxops,
1184                             struct vfs_cred *vcred)
1185 {
1186         struct gss_cli_ctx    *gctx = ctx2gctx(ctx);
1187
1188         gctx->gc_win = 0;
1189         cfs_atomic_set(&gctx->gc_seq, 0);
1190
1191         CFS_INIT_HLIST_NODE(&ctx->cc_cache);
1192         cfs_atomic_set(&ctx->cc_refcount, 0);
1193         ctx->cc_sec = sec;
1194         ctx->cc_ops = ctxops;
1195         ctx->cc_expire = 0;
1196         ctx->cc_flags = PTLRPC_CTX_NEW;
1197         ctx->cc_vcred = *vcred;
1198         cfs_spin_lock_init(&ctx->cc_lock);
1199         CFS_INIT_LIST_HEAD(&ctx->cc_req_list);
1200         CFS_INIT_LIST_HEAD(&ctx->cc_gc_chain);
1201
1202         /* take a ref on belonging sec, balanced in ctx destroying */
1203         cfs_atomic_inc(&sec->ps_refcount);
1204         /* statistic only */
1205         cfs_atomic_inc(&sec->ps_nctx);
1206
1207         CDEBUG(D_SEC, "%s@%p: create ctx %p(%u->%s)\n",
1208                sec->ps_policy->sp_name, ctx->cc_sec,
1209                ctx, ctx->cc_vcred.vc_uid, sec2target_str(ctx->cc_sec));
1210         return 0;
1211 }
1212
1213 /*
1214  * return value:
1215  *   1: the context has been taken care of by someone else
1216  *   0: proceed to really destroy the context locally
1217  */
1218 int gss_cli_ctx_fini_common(struct ptlrpc_sec *sec,
1219                             struct ptlrpc_cli_ctx *ctx)
1220 {
1221         struct gss_cli_ctx *gctx = ctx2gctx(ctx);
1222
1223         LASSERT(cfs_atomic_read(&sec->ps_nctx) > 0);
1224         LASSERT(cfs_atomic_read(&ctx->cc_refcount) == 0);
1225         LASSERT(ctx->cc_sec == sec);
1226
1227         /*
1228          * remove UPTODATE flag of reverse ctx thus we won't send fini rpc,
1229          * this is to avoid potential problems of client side reverse svc ctx
1230          * be mis-destroyed in various recovery senarios. anyway client can
1231          * manage its reverse ctx well by associating it with its buddy ctx.
1232          */
1233         if (sec_is_reverse(sec))
1234                 ctx->cc_flags &= ~PTLRPC_CTX_UPTODATE;
1235
1236         if (gctx->gc_mechctx) {
1237                 /* the final context fini rpc will use this ctx too, and it's
1238                  * asynchronous which finished by request_out_callback(). so
1239                  * we add refcount, whoever drop finally drop the refcount to
1240                  * 0 should responsible for the rest of destroy. */
1241                 cfs_atomic_inc(&ctx->cc_refcount);
1242
1243                 gss_do_ctx_fini_rpc(gctx);
1244                 gss_cli_ctx_finalize(gctx);
1245
1246                 if (!cfs_atomic_dec_and_test(&ctx->cc_refcount))
1247                         return 1;
1248         }
1249
1250         if (sec_is_reverse(sec))
1251                 CWARN("reverse sec %p: destroy ctx %p\n",
1252                       ctx->cc_sec, ctx);
1253         else
1254                 CWARN("%s@%p: destroy ctx %p(%u->%s)\n",
1255                       sec->ps_policy->sp_name, ctx->cc_sec,
1256                       ctx, ctx->cc_vcred.vc_uid, sec2target_str(ctx->cc_sec));
1257
1258         return 0;
1259 }
1260
1261 static
1262 int gss_alloc_reqbuf_intg(struct ptlrpc_sec *sec,
1263                           struct ptlrpc_request *req,
1264                           int svc, int msgsize)
1265 {
1266         int                       bufsize, txtsize;
1267         int                       bufcnt = 2;
1268         __u32                     buflens[5];
1269         ENTRY;
1270
1271         /*
1272          * on-wire data layout:
1273          * - gss header
1274          * - lustre message
1275          * - user descriptor (optional)
1276          * - bulk sec descriptor (optional)
1277          * - signature (optional)
1278          *   - svc == NULL: NULL
1279          *   - svc == AUTH: signature of gss header
1280          *   - svc == INTG: signature of all above
1281          *
1282          * if this is context negotiation, reserver fixed space
1283          * at the last (signature) segment regardless of svc mode.
1284          */
1285
1286         buflens[0] = PTLRPC_GSS_HEADER_SIZE;
1287         txtsize = buflens[0];
1288
1289         buflens[1] = msgsize;
1290         if (svc == SPTLRPC_SVC_INTG)
1291                 txtsize += buflens[1];
1292
1293         if (req->rq_pack_udesc) {
1294                 buflens[bufcnt] = sptlrpc_current_user_desc_size();
1295                 if (svc == SPTLRPC_SVC_INTG)
1296                         txtsize += buflens[bufcnt];
1297                 bufcnt++;
1298         }
1299
1300         if (req->rq_pack_bulk) {
1301                 buflens[bufcnt] = gss_cli_bulk_payload(req->rq_cli_ctx,
1302                                                        &req->rq_flvr,
1303                                                        0, req->rq_bulk_read);
1304                 if (svc == SPTLRPC_SVC_INTG)
1305                         txtsize += buflens[bufcnt];
1306                 bufcnt++;
1307         }
1308
1309         if (req->rq_ctx_init)
1310                 buflens[bufcnt++] = GSS_CTX_INIT_MAX_LEN;
1311         else if (svc != SPTLRPC_SVC_NULL)
1312                 buflens[bufcnt++] = gss_cli_payload(req->rq_cli_ctx, txtsize,0);
1313
1314         bufsize = lustre_msg_size_v2(bufcnt, buflens);
1315
1316         if (!req->rq_reqbuf) {
1317                 bufsize = size_roundup_power2(bufsize);
1318
1319                 OBD_ALLOC(req->rq_reqbuf, bufsize);
1320                 if (!req->rq_reqbuf)
1321                         RETURN(-ENOMEM);
1322
1323                 req->rq_reqbuf_len = bufsize;
1324         } else {
1325                 LASSERT(req->rq_pool);
1326                 LASSERT(req->rq_reqbuf_len >= bufsize);
1327                 memset(req->rq_reqbuf, 0, bufsize);
1328         }
1329
1330         lustre_init_msg_v2(req->rq_reqbuf, bufcnt, buflens, NULL);
1331         req->rq_reqbuf->lm_secflvr = req->rq_flvr.sf_rpc;
1332
1333         req->rq_reqmsg = lustre_msg_buf(req->rq_reqbuf, 1, msgsize);
1334         LASSERT(req->rq_reqmsg);
1335
1336         /* pack user desc here, later we might leave current user's process */
1337         if (req->rq_pack_udesc)
1338                 sptlrpc_pack_user_desc(req->rq_reqbuf, 2);
1339
1340         RETURN(0);
1341 }
1342
1343 static
1344 int gss_alloc_reqbuf_priv(struct ptlrpc_sec *sec,
1345                           struct ptlrpc_request *req,
1346                           int msgsize)
1347 {
1348         __u32                     ibuflens[3], wbuflens[2];
1349         int                       ibufcnt;
1350         int                       clearsize, wiresize;
1351         ENTRY;
1352
1353         LASSERT(req->rq_clrbuf == NULL);
1354         LASSERT(req->rq_clrbuf_len == 0);
1355
1356         /* Inner (clear) buffers
1357          *  - lustre message
1358          *  - user descriptor (optional)
1359          *  - bulk checksum (optional)
1360          */
1361         ibufcnt = 1;
1362         ibuflens[0] = msgsize;
1363
1364         if (req->rq_pack_udesc)
1365                 ibuflens[ibufcnt++] = sptlrpc_current_user_desc_size();
1366         if (req->rq_pack_bulk)
1367                 ibuflens[ibufcnt++] = gss_cli_bulk_payload(req->rq_cli_ctx,
1368                                                            &req->rq_flvr, 0,
1369                                                            req->rq_bulk_read);
1370
1371         clearsize = lustre_msg_size_v2(ibufcnt, ibuflens);
1372         /* to allow append padding during encryption */
1373         clearsize += GSS_MAX_CIPHER_BLOCK;
1374
1375         /* Wrapper (wire) buffers
1376          *  - gss header
1377          *  - cipher text
1378          */
1379         wbuflens[0] = PTLRPC_GSS_HEADER_SIZE;
1380         wbuflens[1] = gss_cli_payload(req->rq_cli_ctx, clearsize, 1);
1381         wiresize = lustre_msg_size_v2(2, wbuflens);
1382
1383         if (req->rq_pool) {
1384                 /* rq_reqbuf is preallocated */
1385                 LASSERT(req->rq_reqbuf);
1386                 LASSERT(req->rq_reqbuf_len >= wiresize);
1387
1388                 memset(req->rq_reqbuf, 0, req->rq_reqbuf_len);
1389
1390                 /* if the pre-allocated buffer is big enough, we just pack
1391                  * both clear buf & request buf in it, to avoid more alloc. */
1392                 if (clearsize + wiresize <= req->rq_reqbuf_len) {
1393                         req->rq_clrbuf =
1394                                 (void *) (((char *) req->rq_reqbuf) + wiresize);
1395                 } else {
1396                         CWARN("pre-allocated buf size %d is not enough for "
1397                               "both clear (%d) and cipher (%d) text, proceed "
1398                               "with extra allocation\n", req->rq_reqbuf_len,
1399                               clearsize, wiresize);
1400                 }
1401         }
1402
1403         if (!req->rq_clrbuf) {
1404                 clearsize = size_roundup_power2(clearsize);
1405
1406                 OBD_ALLOC(req->rq_clrbuf, clearsize);
1407                 if (!req->rq_clrbuf)
1408                         RETURN(-ENOMEM);
1409         }
1410         req->rq_clrbuf_len = clearsize;
1411
1412         lustre_init_msg_v2(req->rq_clrbuf, ibufcnt, ibuflens, NULL);
1413         req->rq_reqmsg = lustre_msg_buf(req->rq_clrbuf, 0, msgsize);
1414
1415         if (req->rq_pack_udesc)
1416                 sptlrpc_pack_user_desc(req->rq_clrbuf, 1);
1417
1418         RETURN(0);
1419 }
1420
1421 /*
1422  * NOTE: any change of request buffer allocation should also consider
1423  * changing enlarge_reqbuf() series functions.
1424  */
1425 int gss_alloc_reqbuf(struct ptlrpc_sec *sec,
1426                      struct ptlrpc_request *req,
1427                      int msgsize)
1428 {
1429         int     svc = SPTLRPC_FLVR_SVC(req->rq_flvr.sf_rpc);
1430
1431         LASSERT(!req->rq_pack_bulk ||
1432                 (req->rq_bulk_read || req->rq_bulk_write));
1433
1434         switch (svc) {
1435         case SPTLRPC_SVC_NULL:
1436         case SPTLRPC_SVC_AUTH:
1437         case SPTLRPC_SVC_INTG:
1438                 return gss_alloc_reqbuf_intg(sec, req, svc, msgsize);
1439         case SPTLRPC_SVC_PRIV:
1440                 return gss_alloc_reqbuf_priv(sec, req, msgsize);
1441         default:
1442                 LASSERTF(0, "bad rpc flavor %x\n", req->rq_flvr.sf_rpc);
1443                 return 0;
1444         }
1445 }
1446
1447 void gss_free_reqbuf(struct ptlrpc_sec *sec,
1448                      struct ptlrpc_request *req)
1449 {
1450         int     privacy;
1451         ENTRY;
1452
1453         LASSERT(!req->rq_pool || req->rq_reqbuf);
1454         privacy = SPTLRPC_FLVR_SVC(req->rq_flvr.sf_rpc) == SPTLRPC_SVC_PRIV;
1455
1456         if (!req->rq_clrbuf)
1457                 goto release_reqbuf;
1458
1459         /* release clear buffer */
1460         LASSERT(privacy);
1461         LASSERT(req->rq_clrbuf_len);
1462
1463         if (req->rq_pool == NULL ||
1464             req->rq_clrbuf < req->rq_reqbuf ||
1465             (char *) req->rq_clrbuf >=
1466             (char *) req->rq_reqbuf + req->rq_reqbuf_len)
1467                 OBD_FREE(req->rq_clrbuf, req->rq_clrbuf_len);
1468
1469         req->rq_clrbuf = NULL;
1470         req->rq_clrbuf_len = 0;
1471
1472 release_reqbuf:
1473         if (!req->rq_pool && req->rq_reqbuf) {
1474                 LASSERT(req->rq_reqbuf_len);
1475
1476                 OBD_FREE(req->rq_reqbuf, req->rq_reqbuf_len);
1477                 req->rq_reqbuf = NULL;
1478                 req->rq_reqbuf_len = 0;
1479         }
1480
1481         EXIT;
1482 }
1483
1484 static int do_alloc_repbuf(struct ptlrpc_request *req, int bufsize)
1485 {
1486         bufsize = size_roundup_power2(bufsize);
1487
1488         OBD_ALLOC(req->rq_repbuf, bufsize);
1489         if (!req->rq_repbuf)
1490                 return -ENOMEM;
1491
1492         req->rq_repbuf_len = bufsize;
1493         return 0;
1494 }
1495
1496 static
1497 int gss_alloc_repbuf_intg(struct ptlrpc_sec *sec,
1498                           struct ptlrpc_request *req,
1499                           int svc, int msgsize)
1500 {
1501         int             txtsize;
1502         __u32           buflens[4];
1503         int             bufcnt = 2;
1504         int             alloc_size;
1505
1506         /*
1507          * on-wire data layout:
1508          * - gss header
1509          * - lustre message
1510          * - bulk sec descriptor (optional)
1511          * - signature (optional)
1512          *   - svc == NULL: NULL
1513          *   - svc == AUTH: signature of gss header
1514          *   - svc == INTG: signature of all above
1515          *
1516          * if this is context negotiation, reserver fixed space
1517          * at the last (signature) segment regardless of svc mode.
1518          */
1519
1520         buflens[0] = PTLRPC_GSS_HEADER_SIZE;
1521         txtsize = buflens[0];
1522
1523         buflens[1] = msgsize;
1524         if (svc == SPTLRPC_SVC_INTG)
1525                 txtsize += buflens[1];
1526
1527         if (req->rq_pack_bulk) {
1528                 buflens[bufcnt] = gss_cli_bulk_payload(req->rq_cli_ctx,
1529                                                        &req->rq_flvr,
1530                                                        1, req->rq_bulk_read);
1531                 if (svc == SPTLRPC_SVC_INTG)
1532                         txtsize += buflens[bufcnt];
1533                 bufcnt++;
1534         }
1535
1536         if (req->rq_ctx_init)
1537                 buflens[bufcnt++] = GSS_CTX_INIT_MAX_LEN;
1538         else if (svc != SPTLRPC_SVC_NULL)
1539                 buflens[bufcnt++] = gss_cli_payload(req->rq_cli_ctx, txtsize,0);
1540
1541         alloc_size = lustre_msg_size_v2(bufcnt, buflens);
1542
1543         /* add space for early reply */
1544         alloc_size += gss_at_reply_off_integ;
1545
1546         return do_alloc_repbuf(req, alloc_size);
1547 }
1548
1549 static
1550 int gss_alloc_repbuf_priv(struct ptlrpc_sec *sec,
1551                           struct ptlrpc_request *req,
1552                           int msgsize)
1553 {
1554         int             txtsize;
1555         __u32           buflens[2];
1556         int             bufcnt;
1557         int             alloc_size;
1558
1559         /* inner buffers */
1560         bufcnt = 1;
1561         buflens[0] = msgsize;
1562
1563         if (req->rq_pack_bulk)
1564                 buflens[bufcnt++] = gss_cli_bulk_payload(req->rq_cli_ctx,
1565                                                          &req->rq_flvr,
1566                                                          1, req->rq_bulk_read);
1567         txtsize = lustre_msg_size_v2(bufcnt, buflens);
1568         txtsize += GSS_MAX_CIPHER_BLOCK;
1569
1570         /* wrapper buffers */
1571         bufcnt = 2;
1572         buflens[0] = PTLRPC_GSS_HEADER_SIZE;
1573         buflens[1] = gss_cli_payload(req->rq_cli_ctx, txtsize, 1);
1574
1575         alloc_size = lustre_msg_size_v2(bufcnt, buflens);
1576         /* add space for early reply */
1577         alloc_size += gss_at_reply_off_priv;
1578
1579         return do_alloc_repbuf(req, alloc_size);
1580 }
1581
1582 int gss_alloc_repbuf(struct ptlrpc_sec *sec,
1583                      struct ptlrpc_request *req,
1584                      int msgsize)
1585 {
1586         int     svc = SPTLRPC_FLVR_SVC(req->rq_flvr.sf_rpc);
1587         ENTRY;
1588
1589         LASSERT(!req->rq_pack_bulk ||
1590                 (req->rq_bulk_read || req->rq_bulk_write));
1591
1592         switch (svc) {
1593         case SPTLRPC_SVC_NULL:
1594         case SPTLRPC_SVC_AUTH:
1595         case SPTLRPC_SVC_INTG:
1596                 return gss_alloc_repbuf_intg(sec, req, svc, msgsize);
1597         case SPTLRPC_SVC_PRIV:
1598                 return gss_alloc_repbuf_priv(sec, req, msgsize);
1599         default:
1600                 LASSERTF(0, "bad rpc flavor %x\n", req->rq_flvr.sf_rpc);
1601                 return 0;
1602         }
1603 }
1604
1605 void gss_free_repbuf(struct ptlrpc_sec *sec,
1606                      struct ptlrpc_request *req)
1607 {
1608         OBD_FREE(req->rq_repbuf, req->rq_repbuf_len);
1609         req->rq_repbuf = NULL;
1610         req->rq_repbuf_len = 0;
1611         req->rq_repdata = NULL;
1612         req->rq_repdata_len = 0;
1613 }
1614
1615 static int get_enlarged_msgsize(struct lustre_msg *msg,
1616                                 int segment, int newsize)
1617 {
1618         int save, newmsg_size;
1619
1620         LASSERT(newsize >= msg->lm_buflens[segment]);
1621
1622         save = msg->lm_buflens[segment];
1623         msg->lm_buflens[segment] = newsize;
1624         newmsg_size = lustre_msg_size_v2(msg->lm_bufcount, msg->lm_buflens);
1625         msg->lm_buflens[segment] = save;
1626
1627         return newmsg_size;
1628 }
1629
1630 static int get_enlarged_msgsize2(struct lustre_msg *msg,
1631                                  int segment1, int newsize1,
1632                                  int segment2, int newsize2)
1633 {
1634         int save1, save2, newmsg_size;
1635
1636         LASSERT(newsize1 >= msg->lm_buflens[segment1]);
1637         LASSERT(newsize2 >= msg->lm_buflens[segment2]);
1638
1639         save1 = msg->lm_buflens[segment1];
1640         save2 = msg->lm_buflens[segment2];
1641         msg->lm_buflens[segment1] = newsize1;
1642         msg->lm_buflens[segment2] = newsize2;
1643         newmsg_size = lustre_msg_size_v2(msg->lm_bufcount, msg->lm_buflens);
1644         msg->lm_buflens[segment1] = save1;
1645         msg->lm_buflens[segment2] = save2;
1646
1647         return newmsg_size;
1648 }
1649
1650 static
1651 int gss_enlarge_reqbuf_intg(struct ptlrpc_sec *sec,
1652                             struct ptlrpc_request *req,
1653                             int svc,
1654                             int segment, int newsize)
1655 {
1656         struct lustre_msg      *newbuf;
1657         int                     txtsize, sigsize = 0, i;
1658         int                     newmsg_size, newbuf_size;
1659
1660         /*
1661          * gss header is at seg 0;
1662          * embedded msg is at seg 1;
1663          * signature (if any) is at the last seg
1664          */
1665         LASSERT(req->rq_reqbuf);
1666         LASSERT(req->rq_reqbuf_len > req->rq_reqlen);
1667         LASSERT(req->rq_reqbuf->lm_bufcount >= 2);
1668         LASSERT(lustre_msg_buf(req->rq_reqbuf, 1, 0) == req->rq_reqmsg);
1669
1670         /* 1. compute new embedded msg size */
1671         newmsg_size = get_enlarged_msgsize(req->rq_reqmsg, segment, newsize);
1672         LASSERT(newmsg_size >= req->rq_reqbuf->lm_buflens[1]);
1673
1674         /* 2. compute new wrapper msg size */
1675         if (svc == SPTLRPC_SVC_NULL) {
1676                 /* no signature, get size directly */
1677                 newbuf_size = get_enlarged_msgsize(req->rq_reqbuf,
1678                                                    1, newmsg_size);
1679         } else {
1680                 txtsize = req->rq_reqbuf->lm_buflens[0];
1681
1682                 if (svc == SPTLRPC_SVC_INTG) {
1683                         for (i = 1; i < req->rq_reqbuf->lm_bufcount; i++)
1684                                 txtsize += req->rq_reqbuf->lm_buflens[i];
1685                         txtsize += newmsg_size - req->rq_reqbuf->lm_buflens[1];
1686                 }
1687
1688                 sigsize = gss_cli_payload(req->rq_cli_ctx, txtsize, 0);
1689                 LASSERT(sigsize >= msg_last_seglen(req->rq_reqbuf));
1690
1691                 newbuf_size = get_enlarged_msgsize2(
1692                                         req->rq_reqbuf,
1693                                         1, newmsg_size,
1694                                         msg_last_segidx(req->rq_reqbuf),
1695                                         sigsize);
1696         }
1697
1698         /* request from pool should always have enough buffer */
1699         LASSERT(!req->rq_pool || req->rq_reqbuf_len >= newbuf_size);
1700
1701         if (req->rq_reqbuf_len < newbuf_size) {
1702                 newbuf_size = size_roundup_power2(newbuf_size);
1703
1704                 OBD_ALLOC(newbuf, newbuf_size);
1705                 if (newbuf == NULL)
1706                         RETURN(-ENOMEM);
1707
1708                 memcpy(newbuf, req->rq_reqbuf, req->rq_reqbuf_len);
1709
1710                 OBD_FREE(req->rq_reqbuf, req->rq_reqbuf_len);
1711                 req->rq_reqbuf = newbuf;
1712                 req->rq_reqbuf_len = newbuf_size;
1713                 req->rq_reqmsg = lustre_msg_buf(req->rq_reqbuf, 1, 0);
1714         }
1715
1716         /* do enlargement, from wrapper to embedded, from end to begin */
1717         if (svc != SPTLRPC_SVC_NULL)
1718                 _sptlrpc_enlarge_msg_inplace(req->rq_reqbuf,
1719                                              msg_last_segidx(req->rq_reqbuf),
1720                                              sigsize);
1721
1722         _sptlrpc_enlarge_msg_inplace(req->rq_reqbuf, 1, newmsg_size);
1723         _sptlrpc_enlarge_msg_inplace(req->rq_reqmsg, segment, newsize);
1724
1725         req->rq_reqlen = newmsg_size;
1726         RETURN(0);
1727 }
1728
1729 static
1730 int gss_enlarge_reqbuf_priv(struct ptlrpc_sec *sec,
1731                             struct ptlrpc_request *req,
1732                             int segment, int newsize)
1733 {
1734         struct lustre_msg      *newclrbuf;
1735         int                     newmsg_size, newclrbuf_size, newcipbuf_size;
1736         __u32                   buflens[3];
1737
1738         /*
1739          * embedded msg is at seg 0 of clear buffer;
1740          * cipher text is at seg 2 of cipher buffer;
1741          */
1742         LASSERT(req->rq_pool ||
1743                 (req->rq_reqbuf == NULL && req->rq_reqbuf_len == 0));
1744         LASSERT(req->rq_reqbuf == NULL ||
1745                 (req->rq_pool && req->rq_reqbuf->lm_bufcount == 3));
1746         LASSERT(req->rq_clrbuf);
1747         LASSERT(req->rq_clrbuf_len > req->rq_reqlen);
1748         LASSERT(lustre_msg_buf(req->rq_clrbuf, 0, 0) == req->rq_reqmsg);
1749
1750         /* compute new embedded msg size */
1751         newmsg_size = get_enlarged_msgsize(req->rq_reqmsg, segment, newsize);
1752
1753         /* compute new clear buffer size */
1754         newclrbuf_size = get_enlarged_msgsize(req->rq_clrbuf, 0, newmsg_size);
1755         newclrbuf_size += GSS_MAX_CIPHER_BLOCK;
1756
1757         /* compute new cipher buffer size */
1758         buflens[0] = PTLRPC_GSS_HEADER_SIZE;
1759         buflens[1] = gss_cli_payload(req->rq_cli_ctx, buflens[0], 0);
1760         buflens[2] = gss_cli_payload(req->rq_cli_ctx, newclrbuf_size, 1);
1761         newcipbuf_size = lustre_msg_size_v2(3, buflens);
1762
1763         /* handle the case that we put both clear buf and cipher buf into
1764          * pre-allocated single buffer. */
1765         if (unlikely(req->rq_pool) &&
1766             req->rq_clrbuf >= req->rq_reqbuf &&
1767             (char *) req->rq_clrbuf <
1768             (char *) req->rq_reqbuf + req->rq_reqbuf_len) {
1769                 /* it couldn't be better we still fit into the
1770                  * pre-allocated buffer. */
1771                 if (newclrbuf_size + newcipbuf_size <= req->rq_reqbuf_len) {
1772                         void *src, *dst;
1773
1774                         /* move clear text backward. */
1775                         src = req->rq_clrbuf;
1776                         dst = (char *) req->rq_reqbuf + newcipbuf_size;
1777
1778                         memmove(dst, src, req->rq_clrbuf_len);
1779
1780                         req->rq_clrbuf = (struct lustre_msg *) dst;
1781                         req->rq_clrbuf_len = newclrbuf_size;
1782                         req->rq_reqmsg = lustre_msg_buf(req->rq_clrbuf, 0, 0);
1783                 } else {
1784                         /* sadly we have to split out the clear buffer */
1785                         LASSERT(req->rq_reqbuf_len >= newcipbuf_size);
1786                         LASSERT(req->rq_clrbuf_len < newclrbuf_size);
1787                 }
1788         }
1789
1790         if (req->rq_clrbuf_len < newclrbuf_size) {
1791                 newclrbuf_size = size_roundup_power2(newclrbuf_size);
1792
1793                 OBD_ALLOC(newclrbuf, newclrbuf_size);
1794                 if (newclrbuf == NULL)
1795                         RETURN(-ENOMEM);
1796
1797                 memcpy(newclrbuf, req->rq_clrbuf, req->rq_clrbuf_len);
1798
1799                 if (req->rq_reqbuf == NULL ||
1800                     req->rq_clrbuf < req->rq_reqbuf ||
1801                     (char *) req->rq_clrbuf >=
1802                     (char *) req->rq_reqbuf + req->rq_reqbuf_len) {
1803                         OBD_FREE(req->rq_clrbuf, req->rq_clrbuf_len);
1804                 }
1805
1806                 req->rq_clrbuf = newclrbuf;
1807                 req->rq_clrbuf_len = newclrbuf_size;
1808                 req->rq_reqmsg = lustre_msg_buf(req->rq_clrbuf, 0, 0);
1809         }
1810
1811         _sptlrpc_enlarge_msg_inplace(req->rq_clrbuf, 0, newmsg_size);
1812         _sptlrpc_enlarge_msg_inplace(req->rq_reqmsg, segment, newsize);
1813         req->rq_reqlen = newmsg_size;
1814
1815         RETURN(0);
1816 }
1817
1818 int gss_enlarge_reqbuf(struct ptlrpc_sec *sec,
1819                        struct ptlrpc_request *req,
1820                        int segment, int newsize)
1821 {
1822         int     svc = SPTLRPC_FLVR_SVC(req->rq_flvr.sf_rpc);
1823
1824         LASSERT(!req->rq_ctx_init && !req->rq_ctx_fini);
1825
1826         switch (svc) {
1827         case SPTLRPC_SVC_NULL:
1828         case SPTLRPC_SVC_AUTH:
1829         case SPTLRPC_SVC_INTG:
1830                 return gss_enlarge_reqbuf_intg(sec, req, svc, segment, newsize);
1831         case SPTLRPC_SVC_PRIV:
1832                 return gss_enlarge_reqbuf_priv(sec, req, segment, newsize);
1833         default:
1834                 LASSERTF(0, "bad rpc flavor %x\n", req->rq_flvr.sf_rpc);
1835                 return 0;
1836         }
1837 }
1838
1839 int gss_sec_install_rctx(struct obd_import *imp,
1840                          struct ptlrpc_sec *sec,
1841                          struct ptlrpc_cli_ctx *ctx)
1842 {
1843         struct gss_sec     *gsec;
1844         struct gss_cli_ctx *gctx;
1845         int                 rc;
1846
1847         gsec = container_of(sec, struct gss_sec, gs_base);
1848         gctx = container_of(ctx, struct gss_cli_ctx, gc_base);
1849
1850         rc = gss_install_rvs_svc_ctx(imp, gsec, gctx);
1851         return rc;
1852 }
1853
1854 /********************************************
1855  * server side API                          *
1856  ********************************************/
1857
1858 static inline
1859 int gss_svc_reqctx_is_special(struct gss_svc_reqctx *grctx)
1860 {
1861         LASSERT(grctx);
1862         return (grctx->src_init || grctx->src_init_continue ||
1863                 grctx->src_err_notify);
1864 }
1865
1866 static
1867 void gss_svc_reqctx_free(struct gss_svc_reqctx *grctx)
1868 {
1869         if (grctx->src_ctx)
1870                 gss_svc_upcall_put_ctx(grctx->src_ctx);
1871
1872         sptlrpc_policy_put(grctx->src_base.sc_policy);
1873         OBD_FREE_PTR(grctx);
1874 }
1875
1876 static inline
1877 void gss_svc_reqctx_addref(struct gss_svc_reqctx *grctx)
1878 {
1879         LASSERT(cfs_atomic_read(&grctx->src_base.sc_refcount) > 0);
1880         cfs_atomic_inc(&grctx->src_base.sc_refcount);
1881 }
1882
1883 static inline
1884 void gss_svc_reqctx_decref(struct gss_svc_reqctx *grctx)
1885 {
1886         LASSERT(cfs_atomic_read(&grctx->src_base.sc_refcount) > 0);
1887
1888         if (cfs_atomic_dec_and_test(&grctx->src_base.sc_refcount))
1889                 gss_svc_reqctx_free(grctx);
1890 }
1891
1892 static
1893 int gss_svc_sign(struct ptlrpc_request *req,
1894                  struct ptlrpc_reply_state *rs,
1895                  struct gss_svc_reqctx *grctx,
1896                  __u32 svc)
1897 {
1898         __u32   flags = 0;
1899         int     rc;
1900         ENTRY;
1901
1902         LASSERT(rs->rs_msg == lustre_msg_buf(rs->rs_repbuf, 1, 0));
1903
1904         /* embedded lustre_msg might have been shrinked */
1905         if (req->rq_replen != rs->rs_repbuf->lm_buflens[1])
1906                 lustre_shrink_msg(rs->rs_repbuf, 1, req->rq_replen, 1);
1907
1908         if (req->rq_pack_bulk)
1909                 flags |= LUSTRE_GSS_PACK_BULK;
1910
1911         rc = gss_sign_msg(rs->rs_repbuf, grctx->src_ctx->gsc_mechctx,
1912                           LUSTRE_SP_ANY, flags, PTLRPC_GSS_PROC_DATA,
1913                           grctx->src_wirectx.gw_seq, svc, NULL);
1914         if (rc < 0)
1915                 RETURN(rc);
1916
1917         rs->rs_repdata_len = rc;
1918
1919         if (likely(req->rq_packed_final)) {
1920                 if (lustre_msghdr_get_flags(req->rq_reqmsg) & MSGHDR_AT_SUPPORT)
1921                         req->rq_reply_off = gss_at_reply_off_integ;
1922                 else
1923                         req->rq_reply_off = 0;
1924         } else {
1925                 if (svc == SPTLRPC_SVC_NULL)
1926                         rs->rs_repbuf->lm_cksum = crc32_le(!(__u32) 0,
1927                                         lustre_msg_buf(rs->rs_repbuf, 1, 0),
1928                                         lustre_msg_buflen(rs->rs_repbuf, 1));
1929                 req->rq_reply_off = 0;
1930         }
1931
1932         RETURN(0);
1933 }
1934
1935 int gss_pack_err_notify(struct ptlrpc_request *req, __u32 major, __u32 minor)
1936 {
1937         struct gss_svc_reqctx     *grctx = gss_svc_ctx2reqctx(req->rq_svc_ctx);
1938         struct ptlrpc_reply_state *rs;
1939         struct gss_err_header     *ghdr;
1940         int                        replen = sizeof(struct ptlrpc_body);
1941         int                        rc;
1942         ENTRY;
1943
1944         //if (OBD_FAIL_CHECK_ORSET(OBD_FAIL_SVCGSS_ERR_NOTIFY, OBD_FAIL_ONCE))
1945         //      RETURN(-EINVAL);
1946
1947         grctx->src_err_notify = 1;
1948         grctx->src_reserve_len = 0;
1949
1950         rc = lustre_pack_reply_v2(req, 1, &replen, NULL, 0);
1951         if (rc) {
1952                 CERROR("could not pack reply, err %d\n", rc);
1953                 RETURN(rc);
1954         }
1955
1956         /* gss hdr */
1957         rs = req->rq_reply_state;
1958         LASSERT(rs->rs_repbuf->lm_buflens[1] >= sizeof(*ghdr));
1959         ghdr = lustre_msg_buf(rs->rs_repbuf, 0, 0);
1960         ghdr->gh_version = PTLRPC_GSS_VERSION;
1961         ghdr->gh_flags = 0;
1962         ghdr->gh_proc = PTLRPC_GSS_PROC_ERR;
1963         ghdr->gh_major = major;
1964         ghdr->gh_minor = minor;
1965         ghdr->gh_handle.len = 0; /* fake context handle */
1966
1967         rs->rs_repdata_len = lustre_msg_size_v2(rs->rs_repbuf->lm_bufcount,
1968                                                 rs->rs_repbuf->lm_buflens);
1969
1970         CDEBUG(D_SEC, "prepare gss error notify(0x%x/0x%x) to %s\n",
1971                major, minor, libcfs_nid2str(req->rq_peer.nid));
1972         RETURN(0);
1973 }
1974
1975 static
1976 int gss_svc_handle_init(struct ptlrpc_request *req,
1977                         struct gss_wire_ctx *gw)
1978 {
1979         struct gss_svc_reqctx     *grctx = gss_svc_ctx2reqctx(req->rq_svc_ctx);
1980         struct lustre_msg         *reqbuf = req->rq_reqbuf;
1981         struct obd_uuid           *uuid;
1982         struct obd_device         *target;
1983         rawobj_t                   uuid_obj, rvs_hdl, in_token;
1984         __u32                      lustre_svc;
1985         __u32                     *secdata, seclen;
1986         int                        swabbed, rc;
1987         ENTRY;
1988
1989         CDEBUG(D_SEC, "processing gss init(%d) request from %s\n", gw->gw_proc,
1990                libcfs_nid2str(req->rq_peer.nid));
1991
1992         req->rq_ctx_init = 1;
1993
1994         if (gw->gw_flags & LUSTRE_GSS_PACK_BULK) {
1995                 CERROR("unexpected bulk flag\n");
1996                 RETURN(SECSVC_DROP);
1997         }
1998
1999         if (gw->gw_proc == PTLRPC_GSS_PROC_INIT && gw->gw_handle.len != 0) {
2000                 CERROR("proc %u: invalid handle length %u\n",
2001                        gw->gw_proc, gw->gw_handle.len);
2002                 RETURN(SECSVC_DROP);
2003         }
2004
2005         if (reqbuf->lm_bufcount < 3 || reqbuf->lm_bufcount > 4){
2006                 CERROR("Invalid bufcount %d\n", reqbuf->lm_bufcount);
2007                 RETURN(SECSVC_DROP);
2008         }
2009
2010         swabbed = ptlrpc_req_need_swab(req);
2011
2012         /* ctx initiate payload is in last segment */
2013         secdata = lustre_msg_buf(reqbuf, reqbuf->lm_bufcount - 1, 0);
2014         seclen = reqbuf->lm_buflens[reqbuf->lm_bufcount - 1];
2015
2016         if (seclen < 4 + 4) {
2017                 CERROR("sec size %d too small\n", seclen);
2018                 RETURN(SECSVC_DROP);
2019         }
2020
2021         /* lustre svc type */
2022         lustre_svc = le32_to_cpu(*secdata++);
2023         seclen -= 4;
2024
2025         /* extract target uuid, note this code is somewhat fragile
2026          * because touched internal structure of obd_uuid */
2027         if (rawobj_extract(&uuid_obj, &secdata, &seclen)) {
2028                 CERROR("failed to extract target uuid\n");
2029                 RETURN(SECSVC_DROP);
2030         }
2031         uuid_obj.data[uuid_obj.len - 1] = '\0';
2032
2033         uuid = (struct obd_uuid *) uuid_obj.data;
2034         target = class_uuid2obd(uuid);
2035         if (!target || target->obd_stopping || !target->obd_set_up) {
2036                 CERROR("target '%s' is not available for context init (%s)\n",
2037                        uuid->uuid, target == NULL ? "no target" :
2038                        (target->obd_stopping ? "stopping" : "not set up"));
2039                 RETURN(SECSVC_DROP);
2040         }
2041
2042         /* extract reverse handle */
2043         if (rawobj_extract(&rvs_hdl, &secdata, &seclen)) {
2044                 CERROR("failed extract reverse handle\n");
2045                 RETURN(SECSVC_DROP);
2046         }
2047
2048         /* extract token */
2049         if (rawobj_extract(&in_token, &secdata, &seclen)) {
2050                 CERROR("can't extract token\n");
2051                 RETURN(SECSVC_DROP);
2052         }
2053
2054         rc = gss_svc_upcall_handle_init(req, grctx, gw, target, lustre_svc,
2055                                         &rvs_hdl, &in_token);
2056         if (rc != SECSVC_OK)
2057                 RETURN(rc);
2058
2059         if (grctx->src_ctx->gsc_usr_mds || grctx->src_ctx->gsc_usr_oss ||
2060             grctx->src_ctx->gsc_usr_root)
2061                 CWARN("create svc ctx %p: user from %s authenticated as %s\n",
2062                       grctx->src_ctx, libcfs_nid2str(req->rq_peer.nid),
2063                       grctx->src_ctx->gsc_usr_mds ? "mds" :
2064                         (grctx->src_ctx->gsc_usr_oss ? "oss" : "root"));
2065         else
2066                 CWARN("create svc ctx %p: accept user %u from %s\n",
2067                       grctx->src_ctx, grctx->src_ctx->gsc_uid,
2068                       libcfs_nid2str(req->rq_peer.nid));
2069
2070         if (gw->gw_flags & LUSTRE_GSS_PACK_USER) {
2071                 if (reqbuf->lm_bufcount < 4) {
2072                         CERROR("missing user descriptor\n");
2073                         RETURN(SECSVC_DROP);
2074                 }
2075                 if (sptlrpc_unpack_user_desc(reqbuf, 2, swabbed)) {
2076                         CERROR("Mal-formed user descriptor\n");
2077                         RETURN(SECSVC_DROP);
2078                 }
2079
2080                 req->rq_pack_udesc = 1;
2081                 req->rq_user_desc = lustre_msg_buf(reqbuf, 2, 0);
2082         }
2083
2084         req->rq_reqmsg = lustre_msg_buf(reqbuf, 1, 0);
2085         req->rq_reqlen = lustre_msg_buflen(reqbuf, 1);
2086
2087         RETURN(rc);
2088 }
2089
2090 /*
2091  * last segment must be the gss signature.
2092  */
2093 static
2094 int gss_svc_verify_request(struct ptlrpc_request *req,
2095                            struct gss_svc_reqctx *grctx,
2096                            struct gss_wire_ctx *gw,
2097                            __u32 *major)
2098 {
2099         struct gss_svc_ctx *gctx = grctx->src_ctx;
2100         struct lustre_msg  *msg = req->rq_reqbuf;
2101         int                 offset = 2;
2102         int                 swabbed;
2103         ENTRY;
2104
2105         *major = GSS_S_COMPLETE;
2106
2107         if (msg->lm_bufcount < 2) {
2108                 CERROR("Too few segments (%u) in request\n", msg->lm_bufcount);
2109                 RETURN(-EINVAL);
2110         }
2111
2112         if (gw->gw_svc == SPTLRPC_SVC_NULL)
2113                 goto verified;
2114
2115         if (gss_check_seq_num(&gctx->gsc_seqdata, gw->gw_seq, 0)) {
2116                 CERROR("phase 0: discard replayed req: seq %u\n", gw->gw_seq);
2117                 *major = GSS_S_DUPLICATE_TOKEN;
2118                 RETURN(-EACCES);
2119         }
2120
2121         *major = gss_verify_msg(msg, gctx->gsc_mechctx, gw->gw_svc);
2122         if (*major != GSS_S_COMPLETE) {
2123                 CERROR("failed to verify request: %x\n", *major);
2124                 RETURN(-EACCES);
2125         }
2126
2127         if (gctx->gsc_reverse == 0 &&
2128             gss_check_seq_num(&gctx->gsc_seqdata, gw->gw_seq, 1)) {
2129                 CERROR("phase 1+: discard replayed req: seq %u\n", gw->gw_seq);
2130                 *major = GSS_S_DUPLICATE_TOKEN;
2131                 RETURN(-EACCES);
2132         }
2133
2134 verified:
2135         swabbed = ptlrpc_req_need_swab(req);
2136
2137         /* user descriptor */
2138         if (gw->gw_flags & LUSTRE_GSS_PACK_USER) {
2139                 if (msg->lm_bufcount < (offset + 1)) {
2140                         CERROR("no user desc included\n");
2141                         RETURN(-EINVAL);
2142                 }
2143
2144                 if (sptlrpc_unpack_user_desc(msg, offset, swabbed)) {
2145                         CERROR("Mal-formed user descriptor\n");
2146                         RETURN(-EINVAL);
2147                 }
2148
2149                 req->rq_pack_udesc = 1;
2150                 req->rq_user_desc = lustre_msg_buf(msg, offset, 0);
2151                 offset++;
2152         }
2153
2154         /* check bulk_sec_desc data */
2155         if (gw->gw_flags & LUSTRE_GSS_PACK_BULK) {
2156                 if (msg->lm_bufcount < (offset + 1)) {
2157                         CERROR("missing bulk sec descriptor\n");
2158                         RETURN(-EINVAL);
2159                 }
2160
2161                 if (bulk_sec_desc_unpack(msg, offset, swabbed))
2162                         RETURN(-EINVAL);
2163
2164                 req->rq_pack_bulk = 1;
2165                 grctx->src_reqbsd = lustre_msg_buf(msg, offset, 0);
2166                 grctx->src_reqbsd_size = lustre_msg_buflen(msg, offset);
2167         }
2168
2169         req->rq_reqmsg = lustre_msg_buf(msg, 1, 0);
2170         req->rq_reqlen = msg->lm_buflens[1];
2171         RETURN(0);
2172 }
2173
2174 static
2175 int gss_svc_unseal_request(struct ptlrpc_request *req,
2176                            struct gss_svc_reqctx *grctx,
2177                            struct gss_wire_ctx *gw,
2178                            __u32 *major)
2179 {
2180         struct gss_svc_ctx *gctx = grctx->src_ctx;
2181         struct lustre_msg  *msg = req->rq_reqbuf;
2182         int                 swabbed, msglen, offset = 1;
2183         ENTRY;
2184
2185         if (gss_check_seq_num(&gctx->gsc_seqdata, gw->gw_seq, 0)) {
2186                 CERROR("phase 0: discard replayed req: seq %u\n", gw->gw_seq);
2187                 *major = GSS_S_DUPLICATE_TOKEN;
2188                 RETURN(-EACCES);
2189         }
2190
2191         *major = gss_unseal_msg(gctx->gsc_mechctx, msg,
2192                                &msglen, req->rq_reqdata_len);
2193         if (*major != GSS_S_COMPLETE) {
2194                 CERROR("failed to unwrap request: %x\n", *major);
2195                 RETURN(-EACCES);
2196         }
2197
2198         if (gss_check_seq_num(&gctx->gsc_seqdata, gw->gw_seq, 1)) {
2199                 CERROR("phase 1+: discard replayed req: seq %u\n", gw->gw_seq);
2200                 *major = GSS_S_DUPLICATE_TOKEN;
2201                 RETURN(-EACCES);
2202         }
2203
2204         swabbed = __lustre_unpack_msg(msg, msglen);
2205         if (swabbed < 0) {
2206                 CERROR("Failed to unpack after decryption\n");
2207                 RETURN(-EINVAL);
2208         }
2209         req->rq_reqdata_len = msglen;
2210
2211         if (msg->lm_bufcount < 1) {
2212                 CERROR("Invalid buffer: is empty\n");
2213                 RETURN(-EINVAL);
2214         }
2215
2216         if (gw->gw_flags & LUSTRE_GSS_PACK_USER) {
2217                 if (msg->lm_bufcount < offset + 1) {
2218                         CERROR("no user descriptor included\n");
2219                         RETURN(-EINVAL);
2220                 }
2221
2222                 if (sptlrpc_unpack_user_desc(msg, offset, swabbed)) {
2223                         CERROR("Mal-formed user descriptor\n");
2224                         RETURN(-EINVAL);
2225                 }
2226
2227                 req->rq_pack_udesc = 1;
2228                 req->rq_user_desc = lustre_msg_buf(msg, offset, 0);
2229                 offset++;
2230         }
2231
2232         if (gw->gw_flags & LUSTRE_GSS_PACK_BULK) {
2233                 if (msg->lm_bufcount < offset + 1) {
2234                         CERROR("no bulk checksum included\n");
2235                         RETURN(-EINVAL);
2236                 }
2237
2238                 if (bulk_sec_desc_unpack(msg, offset, swabbed))
2239                         RETURN(-EINVAL);
2240
2241                 req->rq_pack_bulk = 1;
2242                 grctx->src_reqbsd = lustre_msg_buf(msg, offset, 0);
2243                 grctx->src_reqbsd_size = lustre_msg_buflen(msg, offset);
2244         }
2245
2246         req->rq_reqmsg = lustre_msg_buf(req->rq_reqbuf, 0, 0);
2247         req->rq_reqlen = req->rq_reqbuf->lm_buflens[0];
2248         RETURN(0);
2249 }
2250
2251 static
2252 int gss_svc_handle_data(struct ptlrpc_request *req,
2253                         struct gss_wire_ctx *gw)
2254 {
2255         struct gss_svc_reqctx *grctx = gss_svc_ctx2reqctx(req->rq_svc_ctx);
2256         __u32                  major = 0;
2257         int                    rc = 0;
2258         ENTRY;
2259
2260         grctx->src_ctx = gss_svc_upcall_get_ctx(req, gw);
2261         if (!grctx->src_ctx) {
2262                 major = GSS_S_NO_CONTEXT;
2263                 goto error;
2264         }
2265
2266         switch (gw->gw_svc) {
2267         case SPTLRPC_SVC_NULL:
2268         case SPTLRPC_SVC_AUTH:
2269         case SPTLRPC_SVC_INTG:
2270                 rc = gss_svc_verify_request(req, grctx, gw, &major);
2271                 break;
2272         case SPTLRPC_SVC_PRIV:
2273                 rc = gss_svc_unseal_request(req, grctx, gw, &major);
2274                 break;
2275         default:
2276                 CERROR("unsupported gss service %d\n", gw->gw_svc);
2277                 rc = -EINVAL;
2278         }
2279
2280         if (rc == 0)
2281                 RETURN(SECSVC_OK);
2282
2283         CERROR("svc %u failed: major 0x%08x: req xid "LPU64" ctx %p idx "
2284                LPX64"(%u->%s)\n", gw->gw_svc, major, req->rq_xid,
2285                grctx->src_ctx, gss_handle_to_u64(&gw->gw_handle),
2286                grctx->src_ctx->gsc_uid, libcfs_nid2str(req->rq_peer.nid));
2287 error:
2288         /* we only notify client in case of NO_CONTEXT/BAD_SIG, which
2289          * might happen after server reboot, to allow recovery. */
2290         if ((major == GSS_S_NO_CONTEXT || major == GSS_S_BAD_SIG) &&
2291             gss_pack_err_notify(req, major, 0) == 0)
2292                 RETURN(SECSVC_COMPLETE);
2293
2294         RETURN(SECSVC_DROP);
2295 }
2296
2297 static
2298 int gss_svc_handle_destroy(struct ptlrpc_request *req,
2299                            struct gss_wire_ctx *gw)
2300 {
2301         struct gss_svc_reqctx  *grctx = gss_svc_ctx2reqctx(req->rq_svc_ctx);
2302         __u32                   major;
2303         ENTRY;
2304
2305         req->rq_ctx_fini = 1;
2306         req->rq_no_reply = 1;
2307
2308         grctx->src_ctx = gss_svc_upcall_get_ctx(req, gw);
2309         if (!grctx->src_ctx) {
2310                 CDEBUG(D_SEC, "invalid gss context handle for destroy.\n");
2311                 RETURN(SECSVC_DROP);
2312         }
2313
2314         if (gw->gw_svc != SPTLRPC_SVC_INTG) {
2315                 CERROR("svc %u is not supported in destroy.\n", gw->gw_svc);
2316                 RETURN(SECSVC_DROP);
2317         }
2318
2319         if (gss_svc_verify_request(req, grctx, gw, &major))
2320                 RETURN(SECSVC_DROP);
2321
2322         CWARN("destroy svc ctx %p idx "LPX64" (%u->%s)\n",
2323               grctx->src_ctx, gss_handle_to_u64(&gw->gw_handle),
2324               grctx->src_ctx->gsc_uid, libcfs_nid2str(req->rq_peer.nid));
2325
2326         gss_svc_upcall_destroy_ctx(grctx->src_ctx);
2327
2328         if (gw->gw_flags & LUSTRE_GSS_PACK_USER) {
2329                 if (req->rq_reqbuf->lm_bufcount < 4) {
2330                         CERROR("missing user descriptor, ignore it\n");
2331                         RETURN(SECSVC_OK);
2332                 }
2333                 if (sptlrpc_unpack_user_desc(req->rq_reqbuf, 2,
2334                                              ptlrpc_req_need_swab(req))) {
2335                         CERROR("Mal-formed user descriptor, ignore it\n");
2336                         RETURN(SECSVC_OK);
2337                 }
2338
2339                 req->rq_pack_udesc = 1;
2340                 req->rq_user_desc = lustre_msg_buf(req->rq_reqbuf, 2, 0);
2341         }
2342
2343         RETURN(SECSVC_OK);
2344 }
2345
2346 int gss_svc_accept(struct ptlrpc_sec_policy *policy, struct ptlrpc_request *req)
2347 {
2348         struct gss_header      *ghdr;
2349         struct gss_svc_reqctx  *grctx;
2350         struct gss_wire_ctx    *gw;
2351         int                     swabbed, rc;
2352         ENTRY;
2353
2354         LASSERT(req->rq_reqbuf);
2355         LASSERT(req->rq_svc_ctx == NULL);
2356
2357         if (req->rq_reqbuf->lm_bufcount < 2) {
2358                 CERROR("buf count only %d\n", req->rq_reqbuf->lm_bufcount);
2359                 RETURN(SECSVC_DROP);
2360         }
2361
2362         swabbed = ptlrpc_req_need_swab(req);
2363
2364         ghdr = gss_swab_header(req->rq_reqbuf, 0, swabbed);
2365         if (ghdr == NULL) {
2366                 CERROR("can't decode gss header\n");
2367                 RETURN(SECSVC_DROP);
2368         }
2369
2370         /* sanity checks */
2371         if (ghdr->gh_version != PTLRPC_GSS_VERSION) {
2372                 CERROR("gss version %u, expect %u\n", ghdr->gh_version,
2373                        PTLRPC_GSS_VERSION);
2374                 RETURN(SECSVC_DROP);
2375         }
2376
2377         req->rq_sp_from = ghdr->gh_sp;
2378
2379         /* alloc grctx data */
2380         OBD_ALLOC_PTR(grctx);
2381         if (!grctx)
2382                 RETURN(SECSVC_DROP);
2383
2384         grctx->src_base.sc_policy = sptlrpc_policy_get(policy);
2385         cfs_atomic_set(&grctx->src_base.sc_refcount, 1);
2386         req->rq_svc_ctx = &grctx->src_base;
2387         gw = &grctx->src_wirectx;
2388
2389         /* save wire context */
2390         gw->gw_flags = ghdr->gh_flags;
2391         gw->gw_proc = ghdr->gh_proc;
2392         gw->gw_seq = ghdr->gh_seq;
2393         gw->gw_svc = ghdr->gh_svc;
2394         rawobj_from_netobj(&gw->gw_handle, &ghdr->gh_handle);
2395
2396         /* keep original wire header which subject to checksum verification */
2397         if (swabbed)
2398                 gss_header_swabber(ghdr);
2399
2400         switch(ghdr->gh_proc) {
2401         case PTLRPC_GSS_PROC_INIT:
2402         case PTLRPC_GSS_PROC_CONTINUE_INIT:
2403                 rc = gss_svc_handle_init(req, gw);
2404                 break;
2405         case PTLRPC_GSS_PROC_DATA:
2406                 rc = gss_svc_handle_data(req, gw);
2407                 break;
2408         case PTLRPC_GSS_PROC_DESTROY:
2409                 rc = gss_svc_handle_destroy(req, gw);
2410                 break;
2411         default:
2412                 CERROR("unknown proc %u\n", gw->gw_proc);
2413                 rc = SECSVC_DROP;
2414                 break;
2415         }
2416
2417         switch (rc) {
2418         case SECSVC_OK:
2419                 LASSERT (grctx->src_ctx);
2420
2421                 req->rq_auth_gss = 1;
2422                 req->rq_auth_remote = grctx->src_ctx->gsc_remote;
2423                 req->rq_auth_usr_mdt = grctx->src_ctx->gsc_usr_mds;
2424                 req->rq_auth_usr_ost = grctx->src_ctx->gsc_usr_oss;
2425                 req->rq_auth_usr_root = grctx->src_ctx->gsc_usr_root;
2426                 req->rq_auth_uid = grctx->src_ctx->gsc_uid;
2427                 req->rq_auth_mapped_uid = grctx->src_ctx->gsc_mapped_uid;
2428                 break;
2429         case SECSVC_COMPLETE:
2430                 break;
2431         case SECSVC_DROP:
2432                 gss_svc_reqctx_free(grctx);
2433                 req->rq_svc_ctx = NULL;
2434                 break;
2435         }
2436
2437         RETURN(rc);
2438 }
2439
2440 void gss_svc_invalidate_ctx(struct ptlrpc_svc_ctx *svc_ctx)
2441 {
2442         struct gss_svc_reqctx  *grctx;
2443         ENTRY;
2444
2445         if (svc_ctx == NULL) {
2446                 EXIT;
2447                 return;
2448         }
2449
2450         grctx = gss_svc_ctx2reqctx(svc_ctx);
2451
2452         CWARN("gss svc invalidate ctx %p(%u)\n",
2453               grctx->src_ctx, grctx->src_ctx->gsc_uid);
2454         gss_svc_upcall_destroy_ctx(grctx->src_ctx);
2455
2456         EXIT;
2457 }
2458
2459 static inline
2460 int gss_svc_payload(struct gss_svc_reqctx *grctx, int early,
2461                     int msgsize, int privacy)
2462 {
2463         /* we should treat early reply normally, but which is actually sharing
2464          * the same ctx with original request, so in this case we should
2465          * ignore the special ctx's special flags */
2466         if (early == 0 && gss_svc_reqctx_is_special(grctx))
2467                 return grctx->src_reserve_len;
2468
2469         return gss_mech_payload(NULL, msgsize, privacy);
2470 }
2471
2472 static int gss_svc_bulk_payload(struct gss_svc_ctx *gctx,
2473                                 struct sptlrpc_flavor *flvr,
2474                                 int read)
2475 {
2476         int     payload = sizeof(struct ptlrpc_bulk_sec_desc);
2477
2478         if (read) {
2479                 switch (SPTLRPC_FLVR_BULK_SVC(flvr->sf_rpc)) {
2480                 case SPTLRPC_BULK_SVC_NULL:
2481                         break;
2482                 case SPTLRPC_BULK_SVC_INTG:
2483                         payload += gss_mech_payload(NULL, 0, 0);
2484                         break;
2485                 case SPTLRPC_BULK_SVC_PRIV:
2486                         payload += gss_mech_payload(NULL, 0, 1);
2487                         break;
2488                 case SPTLRPC_BULK_SVC_AUTH:
2489                 default:
2490                         LBUG();
2491                 }
2492         }
2493
2494         return payload;
2495 }
2496
2497 int gss_svc_alloc_rs(struct ptlrpc_request *req, int msglen)
2498 {
2499         struct gss_svc_reqctx       *grctx;
2500         struct ptlrpc_reply_state   *rs;
2501         int                          early, privacy, svc, bsd_off = 0;
2502         __u32                        ibuflens[2], buflens[4];
2503         int                          ibufcnt = 0, bufcnt;
2504         int                          txtsize, wmsg_size, rs_size;
2505         ENTRY;
2506
2507         LASSERT(msglen % 8 == 0);
2508
2509         if (req->rq_pack_bulk && !req->rq_bulk_read && !req->rq_bulk_write) {
2510                 CERROR("client request bulk sec on non-bulk rpc\n");
2511                 RETURN(-EPROTO);
2512         }
2513
2514         svc = SPTLRPC_FLVR_SVC(req->rq_flvr.sf_rpc);
2515         early = (req->rq_packed_final == 0);
2516
2517         grctx = gss_svc_ctx2reqctx(req->rq_svc_ctx);
2518         if (!early && gss_svc_reqctx_is_special(grctx))
2519                 privacy = 0;
2520         else
2521                 privacy = (svc == SPTLRPC_SVC_PRIV);
2522
2523         if (privacy) {
2524                 /* inner clear buffers */
2525                 ibufcnt = 1;
2526                 ibuflens[0] = msglen;
2527
2528                 if (req->rq_pack_bulk) {
2529                         LASSERT(grctx->src_reqbsd);
2530
2531                         bsd_off = ibufcnt;
2532                         ibuflens[ibufcnt++] = gss_svc_bulk_payload(
2533                                                         grctx->src_ctx,
2534                                                         &req->rq_flvr,
2535                                                         req->rq_bulk_read);
2536                 }
2537
2538                 txtsize = lustre_msg_size_v2(ibufcnt, ibuflens);
2539                 txtsize += GSS_MAX_CIPHER_BLOCK;
2540
2541                 /* wrapper buffer */
2542                 bufcnt = 2;
2543                 buflens[0] = PTLRPC_GSS_HEADER_SIZE;
2544                 buflens[1] = gss_svc_payload(grctx, early, txtsize, 1);
2545         } else {
2546                 bufcnt = 2;
2547                 buflens[0] = PTLRPC_GSS_HEADER_SIZE;
2548                 buflens[1] = msglen;
2549
2550                 txtsize = buflens[0];
2551                 if (svc == SPTLRPC_SVC_INTG)
2552                         txtsize += buflens[1];
2553
2554                 if (req->rq_pack_bulk) {
2555                         LASSERT(grctx->src_reqbsd);
2556
2557                         bsd_off = bufcnt;
2558                         buflens[bufcnt] = gss_svc_bulk_payload(
2559                                                         grctx->src_ctx,
2560                                                         &req->rq_flvr,
2561                                                         req->rq_bulk_read);
2562                         if (svc == SPTLRPC_SVC_INTG)
2563                                 txtsize += buflens[bufcnt];
2564                         bufcnt++;
2565                 }
2566
2567                 if ((!early && gss_svc_reqctx_is_special(grctx)) ||
2568                     svc != SPTLRPC_SVC_NULL)
2569                         buflens[bufcnt++] = gss_svc_payload(grctx, early,
2570                                                             txtsize, 0);
2571         }
2572
2573         wmsg_size = lustre_msg_size_v2(bufcnt, buflens);
2574
2575         rs_size = sizeof(*rs) + wmsg_size;
2576         rs = req->rq_reply_state;
2577
2578         if (rs) {
2579                 /* pre-allocated */
2580                 LASSERT(rs->rs_size >= rs_size);
2581         } else {
2582                 OBD_ALLOC(rs, rs_size);
2583                 if (rs == NULL)
2584                         RETURN(-ENOMEM);
2585
2586                 rs->rs_size = rs_size;
2587         }
2588
2589         rs->rs_repbuf = (struct lustre_msg *) (rs + 1);
2590         rs->rs_repbuf_len = wmsg_size;
2591
2592         /* initialize the buffer */
2593         if (privacy) {
2594                 lustre_init_msg_v2(rs->rs_repbuf, ibufcnt, ibuflens, NULL);
2595                 rs->rs_msg = lustre_msg_buf(rs->rs_repbuf, 0, msglen);
2596         } else {
2597                 lustre_init_msg_v2(rs->rs_repbuf, bufcnt, buflens, NULL);
2598                 rs->rs_repbuf->lm_secflvr = req->rq_flvr.sf_rpc;
2599
2600                 rs->rs_msg = lustre_msg_buf(rs->rs_repbuf, 1, 0);
2601         }
2602
2603         if (bsd_off) {
2604                 grctx->src_repbsd = lustre_msg_buf(rs->rs_repbuf, bsd_off, 0);
2605                 grctx->src_repbsd_size = lustre_msg_buflen(rs->rs_repbuf,
2606                                                            bsd_off);
2607         }
2608
2609         gss_svc_reqctx_addref(grctx);
2610         rs->rs_svc_ctx = req->rq_svc_ctx;
2611
2612         LASSERT(rs->rs_msg);
2613         req->rq_reply_state = rs;
2614         RETURN(0);
2615 }
2616
2617 static int gss_svc_seal(struct ptlrpc_request *req,
2618                         struct ptlrpc_reply_state *rs,
2619                         struct gss_svc_reqctx *grctx)
2620 {
2621         struct gss_svc_ctx      *gctx = grctx->src_ctx;
2622         rawobj_t                 hdrobj, msgobj, token;
2623         struct gss_header       *ghdr;
2624         __u8                    *token_buf;
2625         int                      token_buflen; 
2626         __u32                    buflens[2], major;
2627         int                      msglen, rc;
2628         ENTRY;
2629
2630         /* get clear data length. note embedded lustre_msg might
2631          * have been shrinked */
2632         if (req->rq_replen != lustre_msg_buflen(rs->rs_repbuf, 0))
2633                 msglen = lustre_shrink_msg(rs->rs_repbuf, 0, req->rq_replen, 1);
2634         else 
2635                 msglen = lustre_msg_size_v2(rs->rs_repbuf->lm_bufcount,
2636                                             rs->rs_repbuf->lm_buflens);
2637
2638         /* temporarily use tail of buffer to hold gss header data */
2639         LASSERT(msglen + PTLRPC_GSS_HEADER_SIZE <= rs->rs_repbuf_len);
2640         ghdr = (struct gss_header *) ((char *) rs->rs_repbuf +
2641                                 rs->rs_repbuf_len - PTLRPC_GSS_HEADER_SIZE);
2642         ghdr->gh_version = PTLRPC_GSS_VERSION;
2643         ghdr->gh_sp = LUSTRE_SP_ANY;
2644         ghdr->gh_flags = 0;
2645         ghdr->gh_proc = PTLRPC_GSS_PROC_DATA;
2646         ghdr->gh_seq = grctx->src_wirectx.gw_seq;
2647         ghdr->gh_svc = SPTLRPC_SVC_PRIV;
2648         ghdr->gh_handle.len = 0;
2649         if (req->rq_pack_bulk)
2650                 ghdr->gh_flags |= LUSTRE_GSS_PACK_BULK;
2651
2652         /* allocate temporary cipher buffer */
2653         token_buflen = gss_mech_payload(gctx->gsc_mechctx, msglen, 1);
2654         OBD_ALLOC(token_buf, token_buflen);
2655         if (token_buf == NULL)
2656                 RETURN(-ENOMEM);
2657
2658         hdrobj.len = PTLRPC_GSS_HEADER_SIZE;
2659         hdrobj.data = (__u8 *) ghdr;
2660         msgobj.len = msglen;
2661         msgobj.data = (__u8 *) rs->rs_repbuf;
2662         token.len = token_buflen;
2663         token.data = token_buf;
2664
2665         major = lgss_wrap(gctx->gsc_mechctx, &hdrobj, &msgobj,
2666                           rs->rs_repbuf_len - PTLRPC_GSS_HEADER_SIZE, &token);
2667         if (major != GSS_S_COMPLETE) {
2668                 CERROR("wrap message error: %08x\n", major);
2669                 GOTO(out_free, rc = -EPERM);
2670         }
2671         LASSERT(token.len <= token_buflen);
2672
2673         /* we are about to override data at rs->rs_repbuf, nullify pointers
2674          * to which to catch further illegal usage. */
2675         if (req->rq_pack_bulk) {
2676                 grctx->src_repbsd = NULL;
2677                 grctx->src_repbsd_size = 0;
2678         }
2679
2680         /* now fill the actual wire data
2681          * - gss header
2682          * - gss token
2683          */
2684         buflens[0] = PTLRPC_GSS_HEADER_SIZE;
2685         buflens[1] = token.len;
2686
2687         rs->rs_repdata_len = lustre_msg_size_v2(2, buflens);
2688         LASSERT(rs->rs_repdata_len <= rs->rs_repbuf_len);
2689
2690         lustre_init_msg_v2(rs->rs_repbuf, 2, buflens, NULL);
2691         rs->rs_repbuf->lm_secflvr = req->rq_flvr.sf_rpc;
2692
2693         memcpy(lustre_msg_buf(rs->rs_repbuf, 0, 0), ghdr,
2694                PTLRPC_GSS_HEADER_SIZE);
2695         memcpy(lustre_msg_buf(rs->rs_repbuf, 1, 0), token.data, token.len);
2696
2697         /* reply offset */
2698         if (req->rq_packed_final &&
2699             (lustre_msghdr_get_flags(req->rq_reqmsg) & MSGHDR_AT_SUPPORT))
2700                 req->rq_reply_off = gss_at_reply_off_priv;
2701         else
2702                 req->rq_reply_off = 0;
2703
2704         /* to catch upper layer's further access */
2705         rs->rs_msg = NULL;
2706         req->rq_repmsg = NULL;
2707         req->rq_replen = 0;
2708
2709         rc = 0;
2710 out_free:
2711         OBD_FREE(token_buf, token_buflen);
2712         RETURN(rc);
2713 }
2714
2715 int gss_svc_authorize(struct ptlrpc_request *req)
2716 {
2717         struct ptlrpc_reply_state *rs = req->rq_reply_state;
2718         struct gss_svc_reqctx     *grctx = gss_svc_ctx2reqctx(req->rq_svc_ctx);
2719         struct gss_wire_ctx       *gw = &grctx->src_wirectx;
2720         int                        early, rc;
2721         ENTRY;
2722
2723         early = (req->rq_packed_final == 0);
2724
2725         if (!early && gss_svc_reqctx_is_special(grctx)) {
2726                 LASSERT(rs->rs_repdata_len != 0);
2727
2728                 req->rq_reply_off = gss_at_reply_off_integ;
2729                 RETURN(0);
2730         }
2731
2732         /* early reply could happen in many cases */
2733         if (!early &&
2734             gw->gw_proc != PTLRPC_GSS_PROC_DATA &&
2735             gw->gw_proc != PTLRPC_GSS_PROC_DESTROY) {
2736                 CERROR("proc %d not support\n", gw->gw_proc);
2737                 RETURN(-EINVAL);
2738         }
2739
2740         LASSERT(grctx->src_ctx);
2741
2742         switch (gw->gw_svc) {
2743         case SPTLRPC_SVC_NULL:
2744         case SPTLRPC_SVC_AUTH:
2745         case SPTLRPC_SVC_INTG:
2746                 rc = gss_svc_sign(req, rs, grctx, gw->gw_svc);
2747                 break;
2748         case SPTLRPC_SVC_PRIV:
2749                 rc = gss_svc_seal(req, rs, grctx);
2750                 break;
2751         default:
2752                 CERROR("Unknown service %d\n", gw->gw_svc);
2753                 GOTO(out, rc = -EINVAL);
2754         }
2755         rc = 0;
2756
2757 out:
2758         RETURN(rc);
2759 }
2760
2761 void gss_svc_free_rs(struct ptlrpc_reply_state *rs)
2762 {
2763         struct gss_svc_reqctx *grctx;
2764
2765         LASSERT(rs->rs_svc_ctx);
2766         grctx = container_of(rs->rs_svc_ctx, struct gss_svc_reqctx, src_base);
2767
2768         gss_svc_reqctx_decref(grctx);
2769         rs->rs_svc_ctx = NULL;
2770
2771         if (!rs->rs_prealloc)
2772                 OBD_FREE(rs, rs->rs_size);
2773 }
2774
2775 void gss_svc_free_ctx(struct ptlrpc_svc_ctx *ctx)
2776 {
2777         LASSERT(cfs_atomic_read(&ctx->sc_refcount) == 0);
2778         gss_svc_reqctx_free(gss_svc_ctx2reqctx(ctx));
2779 }
2780
2781 int gss_copy_rvc_cli_ctx(struct ptlrpc_cli_ctx *cli_ctx,
2782                          struct ptlrpc_svc_ctx *svc_ctx)
2783 {
2784         struct gss_cli_ctx     *cli_gctx = ctx2gctx(cli_ctx);
2785         struct gss_svc_ctx     *svc_gctx = gss_svc_ctx2gssctx(svc_ctx);
2786         struct gss_ctx         *mechctx = NULL;
2787
2788         LASSERT(cli_gctx);
2789         LASSERT(svc_gctx && svc_gctx->gsc_mechctx);
2790
2791         cli_gctx->gc_proc = PTLRPC_GSS_PROC_DATA;
2792         cli_gctx->gc_win = GSS_SEQ_WIN;
2793
2794         /* The problem is the reverse ctx might get lost in some recovery
2795          * situations, and the same svc_ctx will be used to re-create it.
2796          * if there's callback be sentout before that, new reverse ctx start
2797          * with sequence 0 will lead to future callback rpc be treated as
2798          * replay.
2799          *
2800          * each reverse root ctx will record its latest sequence number on its
2801          * buddy svcctx before be destroied, so here we continue use it.
2802          */
2803         cfs_atomic_set(&cli_gctx->gc_seq, svc_gctx->gsc_rvs_seq);
2804
2805         if (gss_svc_upcall_dup_handle(&cli_gctx->gc_svc_handle, svc_gctx)) {
2806                 CERROR("failed to dup svc handle\n");
2807                 goto err_out;
2808         }
2809
2810         if (lgss_copy_reverse_context(svc_gctx->gsc_mechctx, &mechctx) !=
2811             GSS_S_COMPLETE) {
2812                 CERROR("failed to copy mech context\n");
2813                 goto err_svc_handle;
2814         }
2815
2816         if (rawobj_dup(&cli_gctx->gc_handle, &svc_gctx->gsc_rvs_hdl)) {
2817                 CERROR("failed to dup reverse handle\n");
2818                 goto err_ctx;
2819         }
2820
2821         cli_gctx->gc_mechctx = mechctx;
2822         gss_cli_ctx_uptodate(cli_gctx);
2823
2824         return 0;
2825
2826 err_ctx:
2827         lgss_delete_sec_context(&mechctx);
2828 err_svc_handle:
2829         rawobj_free(&cli_gctx->gc_svc_handle);
2830 err_out:
2831         return -ENOMEM;
2832 }
2833
2834 static void gss_init_at_reply_offset(void)
2835 {
2836         __u32 buflens[3];
2837         int clearsize;
2838
2839         buflens[0] = PTLRPC_GSS_HEADER_SIZE;
2840         buflens[1] = lustre_msg_early_size();
2841         buflens[2] = gss_cli_payload(NULL, buflens[1], 0);
2842         gss_at_reply_off_integ = lustre_msg_size_v2(3, buflens);
2843
2844         buflens[0] = lustre_msg_early_size();
2845         clearsize = lustre_msg_size_v2(1, buflens);
2846         buflens[0] = PTLRPC_GSS_HEADER_SIZE;
2847         buflens[1] = gss_cli_payload(NULL, clearsize, 0);
2848         buflens[2] = gss_cli_payload(NULL, clearsize, 1);
2849         gss_at_reply_off_priv = lustre_msg_size_v2(3, buflens);
2850 }
2851
2852 int __init sptlrpc_gss_init(void)
2853 {
2854         int rc;
2855
2856         rc = gss_init_lproc();
2857         if (rc)
2858                 return rc;
2859
2860         rc = gss_init_cli_upcall();
2861         if (rc)
2862                 goto out_lproc;
2863
2864         rc = gss_init_svc_upcall();
2865         if (rc)
2866                 goto out_cli_upcall;
2867
2868         rc = init_kerberos_module();
2869         if (rc)
2870                 goto out_svc_upcall;
2871
2872         /* register policy after all other stuff be intialized, because it
2873          * might be in used immediately after the registration. */
2874
2875         rc = gss_init_keyring();
2876         if (rc)
2877                 goto out_kerberos;
2878
2879 #ifdef HAVE_GSS_PIPEFS
2880         rc = gss_init_pipefs();
2881         if (rc)
2882                 goto out_keyring;
2883 #endif
2884
2885         gss_init_at_reply_offset();
2886
2887         return 0;
2888
2889 #ifdef HAVE_GSS_PIPEFS
2890 out_keyring:
2891         gss_exit_keyring();
2892 #endif
2893
2894 out_kerberos:
2895         cleanup_kerberos_module();
2896 out_svc_upcall:
2897         gss_exit_svc_upcall();
2898 out_cli_upcall:
2899         gss_exit_cli_upcall();
2900 out_lproc:
2901         gss_exit_lproc();
2902         return rc;
2903 }
2904
2905 static void __exit sptlrpc_gss_exit(void)
2906 {
2907         gss_exit_keyring();
2908 #ifdef HAVE_GSS_PIPEFS
2909         gss_exit_pipefs();
2910 #endif
2911         cleanup_kerberos_module();
2912         gss_exit_svc_upcall();
2913         gss_exit_cli_upcall();
2914         gss_exit_lproc();
2915 }
2916
2917 MODULE_AUTHOR("Sun Microsystems, Inc. <http://www.lustre.org/>");
2918 MODULE_DESCRIPTION("GSS security policy for Lustre");
2919 MODULE_LICENSE("GPL");
2920
2921 module_init(sptlrpc_gss_init);
2922 module_exit(sptlrpc_gss_exit);