1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
2 * vim:expandtab:shiftwidth=8:tabstop=8:
4 * Copyright (C) 2007 Cluster File Systems, Inc.
5 * Author: Eric Mei <ericm@clusterfs.com>
7 * This file is part of Lustre, http://www.lustre.org.
9 * Lustre is free software; you can redistribute it and/or
10 * modify it under the terms of version 2 of the GNU General Public
11 * License as published by the Free Software Foundation.
13 * Lustre is distributed in the hope that it will be useful,
14 * but WITHOUT ANY WARRANTY; without even the implied warranty of
15 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
16 * GNU General Public License for more details.
18 * You should have received a copy of the GNU General Public License
19 * along with Lustre; if not, write to the Free Software
20 * Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
24 # define EXPORT_SYMTAB
26 #define DEBUG_SUBSYSTEM S_SEC
28 #include <linux/init.h>
29 #include <linux/module.h>
30 #include <linux/slab.h>
31 #include <linux/dcache.h>
33 #include <linux/random.h>
34 #include <linux/crypto.h>
35 #include <linux/key.h>
36 #include <linux/keyctl.h>
37 #include <linux/mutex.h>
38 #include <asm/atomic.h>
40 #include <liblustre.h>
44 #include <obd_class.h>
45 #include <obd_support.h>
46 #include <lustre/lustre_idl.h>
47 #include <lustre_sec.h>
48 #include <lustre_net.h>
49 #include <lustre_import.h>
52 #include "gss_internal.h"
55 static struct ptlrpc_sec_policy gss_policy_keyring;
56 static struct ptlrpc_ctx_ops gss_keyring_ctxops;
57 static struct key_type gss_key_type;
59 static int sec_install_rctx_kr(struct ptlrpc_sec *sec,
60 struct ptlrpc_svc_ctx *svc_ctx);
63 #define task_aux(tsk) (tsk)
67 * the timeout is only for the case that upcall child process die abnormally.
68 * in any other cases it should finally update kernel key.
70 * FIXME we'd better to incorporate the client & server side upcall timeouts
71 * into the framework of Adaptive Timeouts, but we need to figure out how to
72 * make sure that kernel knows the upcall processes is in-progress or died
75 #define KEYRING_UPCALL_TIMEOUT (obd_timeout + obd_timeout)
77 /****************************************
79 ****************************************/
81 #define DUMP_PROCESS_KEYRINGS(tsk) \
83 CWARN("DUMP PK: %s[%u,%u/%u](<-%s[%u,%u/%u]): " \
84 "a %d, t %d, p %d, s %d, u %d, us %d, df %d\n", \
85 tsk->comm, tsk->pid, tsk->uid, tsk->fsuid, \
86 tsk->parent->comm, tsk->parent->pid, \
87 tsk->parent->uid, tsk->parent->fsuid, \
88 task_aux(tsk)->request_key_auth ? \
89 task_aux(tsk)->request_key_auth->serial : 0, \
90 task_aux(tsk)->thread_keyring ? \
91 task_aux(tsk)->thread_keyring->serial : 0, \
92 tsk->signal->process_keyring ? \
93 tsk->signal->process_keyring->serial : 0, \
94 tsk->signal->session_keyring ? \
95 tsk->signal->session_keyring->serial : 0, \
96 tsk->user->uid_keyring ? \
97 tsk->user->uid_keyring->serial : 0, \
98 tsk->user->session_keyring ? \
99 tsk->user->session_keyring->serial : 0, \
100 task_aux(tsk)->jit_keyring \
104 #define DUMP_KEY(key) \
106 CWARN("DUMP KEY: %p(%d) ref %d u%u/g%u desc %s\n", \
107 key, key->serial, atomic_read(&key->usage), \
108 key->uid, key->gid, \
109 key->description ? key->description : "n/a" \
114 static inline void keyring_upcall_lock(struct gss_sec_keyring *gsec_kr)
116 #ifdef HAVE_KEYRING_UPCALL_SERIALIZED
117 mutex_lock(&gsec_kr->gsk_uc_lock);
121 static inline void keyring_upcall_unlock(struct gss_sec_keyring *gsec_kr)
123 #ifdef HAVE_KEYRING_UPCALL_SERIALIZED
124 mutex_unlock(&gsec_kr->gsk_uc_lock);
128 static inline void key_revoke_locked(struct key *key)
130 set_bit(KEY_FLAG_REVOKED, &key->flags);
133 static void ctx_upcall_timeout_kr(unsigned long data)
135 struct ptlrpc_cli_ctx *ctx = (struct ptlrpc_cli_ctx *) data;
136 struct key *key = ctx2gctx_keyring(ctx)->gck_key;
138 CWARN("ctx %p, key %p\n", ctx, key);
143 key_revoke_locked(key);
144 sptlrpc_cli_ctx_wakeup(ctx);
148 void ctx_start_timer_kr(struct ptlrpc_cli_ctx *ctx, long timeout)
150 struct gss_cli_ctx_keyring *gctx_kr = ctx2gctx_keyring(ctx);
151 struct timer_list *timer = gctx_kr->gck_timer;
155 CDEBUG(D_SEC, "ctx %p: start timer %lds\n", ctx, timeout);
156 timeout = timeout * HZ + cfs_time_current();
159 timer->expires = timeout;
160 timer->data = (unsigned long ) ctx;
161 timer->function = ctx_upcall_timeout_kr;
167 * caller should make sure no race with other threads
170 void ctx_clear_timer_kr(struct ptlrpc_cli_ctx *ctx)
172 struct gss_cli_ctx_keyring *gctx_kr = ctx2gctx_keyring(ctx);
173 struct timer_list *timer = gctx_kr->gck_timer;
178 CDEBUG(D_SEC, "ctx %p, key %p\n", ctx, gctx_kr->gck_key);
180 gctx_kr->gck_timer = NULL;
182 del_singleshot_timer_sync(timer);
188 struct ptlrpc_cli_ctx *ctx_create_kr(struct ptlrpc_sec *sec,
189 struct vfs_cred *vcred)
191 struct ptlrpc_cli_ctx *ctx;
192 struct gss_cli_ctx_keyring *gctx_kr;
194 OBD_ALLOC_PTR(gctx_kr);
198 OBD_ALLOC_PTR(gctx_kr->gck_timer);
199 if (gctx_kr->gck_timer == NULL) {
200 OBD_FREE_PTR(gctx_kr);
203 init_timer(gctx_kr->gck_timer);
205 ctx = &gctx_kr->gck_base.gc_base;
207 if (gss_cli_ctx_init_common(sec, ctx, &gss_keyring_ctxops, vcred)) {
208 OBD_FREE_PTR(gctx_kr->gck_timer);
209 OBD_FREE_PTR(gctx_kr);
213 ctx->cc_expire = cfs_time_current_sec() + KEYRING_UPCALL_TIMEOUT;
214 clear_bit(PTLRPC_CTX_NEW_BIT, &ctx->cc_flags);
215 atomic_inc(&ctx->cc_refcount); /* for the caller */
220 static void ctx_destroy_kr(struct ptlrpc_cli_ctx *ctx)
222 struct ptlrpc_sec *sec = ctx->cc_sec;
223 struct gss_cli_ctx_keyring *gctx_kr = ctx2gctx_keyring(ctx);
225 CDEBUG(D_SEC, "destroying ctx %p\n", ctx);
227 /* at this time the association with key has been broken. */
229 LASSERT(atomic_read(&sec->ps_refcount) > 0);
230 LASSERT(atomic_read(&sec->ps_nctx) > 0);
231 LASSERT(test_bit(PTLRPC_CTX_CACHED_BIT, &ctx->cc_flags) == 0);
232 LASSERT(gctx_kr->gck_key == NULL);
234 ctx_clear_timer_kr(ctx);
235 LASSERT(gctx_kr->gck_timer == NULL);
237 if (gss_cli_ctx_fini_common(sec, ctx))
240 OBD_FREE_PTR(gctx_kr);
242 atomic_dec(&sec->ps_nctx);
243 sptlrpc_sec_put(sec);
246 static void ctx_release_kr(struct ptlrpc_cli_ctx *ctx, int sync)
251 atomic_inc(&ctx->cc_refcount);
252 sptlrpc_gc_add_ctx(ctx);
256 static void ctx_put_kr(struct ptlrpc_cli_ctx *ctx, int sync)
258 LASSERT(atomic_read(&ctx->cc_refcount) > 0);
260 if (atomic_dec_and_test(&ctx->cc_refcount))
261 ctx_release_kr(ctx, sync);
265 * key <-> ctx association and rules:
266 * - ctx might not bind with any key
267 * - key/ctx binding is protected by key semaphore (if the key present)
268 * - key and ctx each take a reference of the other
269 * - ctx enlist/unlist is protected by ctx spinlock
270 * - never enlist a ctx after it's been unlisted
271 * - whoever do enlist should also do bind, lock key before enlist:
272 * - lock key -> lock ctx -> enlist -> unlock ctx -> bind -> unlock key
273 * - whoever do unlist should also do unbind:
274 * - lock key -> lock ctx -> unlist -> unlock ctx -> unbind -> unlock key
275 * - lock ctx -> unlist -> unlock ctx -> lock key -> unbind -> unlock key
278 static inline void spin_lock_if(spinlock_t *lock, int condition)
284 static inline void spin_unlock_if(spinlock_t *lock, int condition)
290 static void ctx_enlist_kr(struct ptlrpc_cli_ctx *ctx, int is_root, int locked)
292 struct ptlrpc_sec *sec = ctx->cc_sec;
293 struct gss_sec_keyring *gsec_kr = sec2gsec_keyring(sec);
295 LASSERT(!test_bit(PTLRPC_CTX_CACHED_BIT, &ctx->cc_flags));
296 LASSERT(atomic_read(&ctx->cc_refcount) > 0);
298 spin_lock_if(&sec->ps_lock, !locked);
300 atomic_inc(&ctx->cc_refcount);
301 set_bit(PTLRPC_CTX_CACHED_BIT, &ctx->cc_flags);
302 hlist_add_head(&ctx->cc_cache, &gsec_kr->gsk_clist);
304 gsec_kr->gsk_root_ctx = ctx;
306 spin_unlock_if(&sec->ps_lock, !locked);
310 * Note after this get called, caller should not access ctx again because
311 * it might have been freed, unless caller hold at least one refcount of
314 * return non-zero if we indeed unlist this ctx.
316 static int ctx_unlist_kr(struct ptlrpc_cli_ctx *ctx, int locked)
318 struct ptlrpc_sec *sec = ctx->cc_sec;
319 struct gss_sec_keyring *gsec_kr = sec2gsec_keyring(sec);
321 /* if hashed bit has gone, leave the job to somebody who is doing it */
322 if (test_and_clear_bit(PTLRPC_CTX_CACHED_BIT, &ctx->cc_flags) == 0)
325 /* drop ref inside spin lock to prevent race with other operations */
326 spin_lock_if(&sec->ps_lock, !locked);
328 if (gsec_kr->gsk_root_ctx == ctx)
329 gsec_kr->gsk_root_ctx = NULL;
330 hlist_del_init(&ctx->cc_cache);
331 atomic_dec(&ctx->cc_refcount);
333 spin_unlock_if(&sec->ps_lock, !locked);
339 * bind a key with a ctx together.
340 * caller must hold write lock of the key, as well as ref on key & ctx.
342 static void bind_key_ctx(struct key *key, struct ptlrpc_cli_ctx *ctx)
344 LASSERT(atomic_read(&ctx->cc_refcount) > 0);
345 LASSERT(atomic_read(&key->usage) > 0);
346 LASSERT(ctx2gctx_keyring(ctx)->gck_key == NULL);
347 LASSERT(key->payload.data == NULL);
349 /* at this time context may or may not in list. */
351 atomic_inc(&ctx->cc_refcount);
352 ctx2gctx_keyring(ctx)->gck_key = key;
353 key->payload.data = ctx;
357 * unbind a key and a ctx.
358 * caller must hold write lock, as well as a ref of the key.
360 static void unbind_key_ctx(struct key *key, struct ptlrpc_cli_ctx *ctx)
362 LASSERT(key->payload.data == ctx);
363 LASSERT(test_bit(PTLRPC_CTX_CACHED_BIT, &ctx->cc_flags) == 0);
365 /* must revoke the key, or others may treat it as newly created */
366 key_revoke_locked(key);
368 key->payload.data = NULL;
369 ctx2gctx_keyring(ctx)->gck_key = NULL;
371 /* once ctx get split from key, the timer is meaningless */
372 ctx_clear_timer_kr(ctx);
379 * given a ctx, unbind with its coupled key, if any.
380 * unbind could only be called once, so we don't worry the key be released
383 static void unbind_ctx_kr(struct ptlrpc_cli_ctx *ctx)
385 struct key *key = ctx2gctx_keyring(ctx)->gck_key;
388 LASSERT(key->payload.data == ctx);
391 down_write(&key->sem);
392 unbind_key_ctx(key, ctx);
399 * given a key, unbind with its coupled ctx, if any.
400 * caller must hold write lock, as well as a ref of the key.
402 static void unbind_key_locked(struct key *key)
404 struct ptlrpc_cli_ctx *ctx = key->payload.data;
407 unbind_key_ctx(key, ctx);
411 * unlist a ctx, and unbind from coupled key
413 static void kill_ctx_kr(struct ptlrpc_cli_ctx *ctx)
415 if (ctx_unlist_kr(ctx, 0))
420 * given a key, unlist and unbind with the coupled ctx (if any).
421 * caller must hold write lock, as well as a ref of the key.
423 static void kill_key_locked(struct key *key)
425 struct ptlrpc_cli_ctx *ctx = key->payload.data;
427 if (ctx && ctx_unlist_kr(ctx, 0))
428 unbind_key_locked(key);
432 * caller should hold one ref on contexts in freelist.
434 static void dispose_ctx_list_kr(struct hlist_head *freelist)
436 struct hlist_node *pos, *next;
437 struct ptlrpc_cli_ctx *ctx;
438 struct gss_cli_ctx *gctx;
440 hlist_for_each_entry_safe(ctx, pos, next, freelist, cc_cache) {
441 hlist_del_init(&ctx->cc_cache);
443 /* reverse ctx: update current seq to buddy svcctx if exist.
444 * ideally this should be done at gss_cli_ctx_finalize(), but
445 * the ctx destroy could be delayed by:
446 * 1) ctx still has reference;
447 * 2) ctx destroy is asynchronous;
448 * and reverse import call inval_all_ctx() require this be done
449 *_immediately_ otherwise newly created reverse ctx might copy
450 * the very old sequence number from svcctx. */
451 gctx = ctx2gctx(ctx);
452 if (!rawobj_empty(&gctx->gc_svc_handle) &&
453 sec_is_reverse(gctx->gc_base.cc_sec)) {
454 gss_svc_upcall_update_sequence(&gctx->gc_svc_handle,
455 (__u32) atomic_read(&gctx->gc_seq));
458 /* we need to wakeup waiting reqs here. the context might
459 * be forced released before upcall finished, then the
460 * late-arrived downcall can't find the ctx even. */
461 sptlrpc_cli_ctx_wakeup(ctx);
469 * lookup a root context directly in a sec, return root ctx with a
470 * reference taken or NULL.
473 struct ptlrpc_cli_ctx * sec_lookup_root_ctx_kr(struct ptlrpc_sec *sec)
475 struct gss_sec_keyring *gsec_kr = sec2gsec_keyring(sec);
476 struct ptlrpc_cli_ctx *ctx = NULL;
478 spin_lock(&sec->ps_lock);
480 ctx = gsec_kr->gsk_root_ctx;
482 if (ctx == NULL && unlikely(sec_is_reverse(sec))) {
483 struct hlist_node *node;
484 struct ptlrpc_cli_ctx *tmp;
486 /* reverse ctx, search root ctx in list, choose the one
487 * with shortest expire time, which is most possibly have
488 * an established peer ctx at client side. */
489 hlist_for_each_entry(tmp, node, &gsec_kr->gsk_clist, cc_cache) {
490 if (ctx == NULL || ctx->cc_expire == 0 ||
491 ctx->cc_expire > tmp->cc_expire) {
493 /* promote to be root_ctx */
494 gsec_kr->gsk_root_ctx = ctx;
500 LASSERT(atomic_read(&ctx->cc_refcount) > 0);
501 LASSERT(!hlist_empty(&gsec_kr->gsk_clist));
502 atomic_inc(&ctx->cc_refcount);
505 spin_unlock(&sec->ps_lock);
510 #define RVS_CTX_EXPIRE_NICE (10)
513 void rvs_sec_install_root_ctx_kr(struct ptlrpc_sec *sec,
514 struct ptlrpc_cli_ctx *new_ctx,
517 struct gss_sec_keyring *gsec_kr = sec2gsec_keyring(sec);
518 struct hlist_node *hnode;
519 struct ptlrpc_cli_ctx *ctx;
523 LASSERT(sec_is_reverse(sec));
525 spin_lock(&sec->ps_lock);
527 now = cfs_time_current_sec();
529 /* set all existing ctxs short expiry */
530 hlist_for_each_entry(ctx, hnode, &gsec_kr->gsk_clist, cc_cache) {
531 if (ctx->cc_expire > now + RVS_CTX_EXPIRE_NICE) {
532 ctx->cc_early_expire = 1;
533 ctx->cc_expire = now + RVS_CTX_EXPIRE_NICE;
537 /* if there's root_ctx there, instead obsolete the current
538 * immediately, we leave it continue operating for a little while.
539 * hopefully when the first backward rpc with newest ctx send out,
540 * the client side already have the peer ctx well established. */
541 ctx_enlist_kr(new_ctx, gsec_kr->gsk_root_ctx ? 0 : 1, 1);
544 bind_key_ctx(key, new_ctx);
546 spin_unlock(&sec->ps_lock);
549 static void construct_key_desc(void *buf, int bufsize,
550 struct ptlrpc_sec *sec, uid_t uid)
552 snprintf(buf, bufsize, "%d@%x", uid, sec->ps_id);
553 ((char *)buf)[bufsize - 1] = '\0';
556 /****************************************
558 ****************************************/
561 struct ptlrpc_sec * gss_sec_create_kr(struct obd_import *imp,
562 struct ptlrpc_svc_ctx *svcctx,
563 struct sptlrpc_flavor *sf)
565 struct gss_sec_keyring *gsec_kr;
568 OBD_ALLOC(gsec_kr, sizeof(*gsec_kr));
572 CFS_INIT_HLIST_HEAD(&gsec_kr->gsk_clist);
573 gsec_kr->gsk_root_ctx = NULL;
574 mutex_init(&gsec_kr->gsk_root_uc_lock);
575 #ifdef HAVE_KEYRING_UPCALL_SERIALIZED
576 mutex_init(&gsec_kr->gsk_uc_lock);
579 if (gss_sec_create_common(&gsec_kr->gsk_base, &gss_policy_keyring,
583 if (svcctx != NULL &&
584 sec_install_rctx_kr(&gsec_kr->gsk_base.gs_base, svcctx)) {
585 gss_sec_destroy_common(&gsec_kr->gsk_base);
589 RETURN(&gsec_kr->gsk_base.gs_base);
592 OBD_FREE(gsec_kr, sizeof(*gsec_kr));
597 void gss_sec_destroy_kr(struct ptlrpc_sec *sec)
599 struct gss_sec *gsec = sec2gsec(sec);
600 struct gss_sec_keyring *gsec_kr = sec2gsec_keyring(sec);
602 CDEBUG(D_SEC, "destroy %s@%p\n", sec->ps_policy->sp_name, sec);
604 LASSERT(hlist_empty(&gsec_kr->gsk_clist));
605 LASSERT(gsec_kr->gsk_root_ctx == NULL);
607 gss_sec_destroy_common(gsec);
609 OBD_FREE(gsec_kr, sizeof(*gsec_kr));
612 static inline int user_is_root(struct ptlrpc_sec *sec, struct vfs_cred *vcred)
614 /* except the ROOTONLY flag, treat it as root user only if real uid
615 * is 0, euid/fsuid being 0 are handled as setuid scenarios */
616 if (sec_is_rootonly(sec) || (vcred->vc_uid == 0))
623 * unlink request key from it's ring, which is linked during request_key().
624 * sadly, we have to 'guess' which keyring it's linked to.
626 * FIXME this code is fragile, depend on how request_key_link() is implemented.
628 static void request_key_unlink(struct key *key)
630 struct task_struct *tsk = current;
633 switch (task_aux(tsk)->jit_keyring) {
634 case KEY_REQKEY_DEFL_DEFAULT:
635 case KEY_REQKEY_DEFL_THREAD_KEYRING:
636 ring = key_get(task_aux(tsk)->thread_keyring);
639 case KEY_REQKEY_DEFL_PROCESS_KEYRING:
640 ring = key_get(tsk->signal->process_keyring);
643 case KEY_REQKEY_DEFL_SESSION_KEYRING:
645 ring = key_get(rcu_dereference(tsk->signal->session_keyring));
649 case KEY_REQKEY_DEFL_USER_SESSION_KEYRING:
650 ring = key_get(tsk->user->session_keyring);
652 case KEY_REQKEY_DEFL_USER_KEYRING:
653 ring = key_get(tsk->user->uid_keyring);
655 case KEY_REQKEY_DEFL_GROUP_KEYRING:
661 key_unlink(ring, key);
666 struct ptlrpc_cli_ctx * gss_sec_lookup_ctx_kr(struct ptlrpc_sec *sec,
667 struct vfs_cred *vcred,
668 int create, int remove_dead)
670 struct obd_import *imp = sec->ps_import;
671 struct gss_sec_keyring *gsec_kr = sec2gsec_keyring(sec);
672 struct ptlrpc_cli_ctx *ctx = NULL;
673 unsigned int is_root = 0, create_new = 0;
681 LASSERT(imp != NULL);
683 is_root = user_is_root(sec, vcred);
685 /* a little bit optimization for root context */
687 ctx = sec_lookup_root_ctx_kr(sec);
689 * Only lookup directly for REVERSE sec, which should
692 if (ctx || sec_is_reverse(sec))
696 LASSERT(create != 0);
698 /* for root context, obtain lock and check again, this time hold
699 * the root upcall lock, make sure nobody else populated new root
700 * context after last check. */
702 mutex_lock(&gsec_kr->gsk_root_uc_lock);
704 ctx = sec_lookup_root_ctx_kr(sec);
708 /* update reverse handle for root user */
709 sec2gsec(sec)->gs_rvs_hdl = gss_get_next_ctx_index();
714 /* in case of setuid, key will be constructed as owner of fsuid/fsgid,
715 * but we do authentication based on real uid/gid. the key permission
716 * bits will be exactly as POS_ALL, so only processes who subscribed
717 * this key could have the access, although the quota might be counted
718 * on others (fsuid/fsgid).
720 * keyring will use fsuid/fsgid as upcall parameters, so we have to
721 * encode real uid/gid into callout info.
724 construct_key_desc(desc, sizeof(desc), sec, vcred->vc_uid);
726 /* callout info format:
727 * secid:mech:uid:gid:flags:svc_type:peer_nid:target_uuid
729 coinfo_size = sizeof(struct obd_uuid) + MAX_OBD_NAME + 64;
730 OBD_ALLOC(coinfo, coinfo_size);
734 snprintf(coinfo, coinfo_size, "%d:%s:%u:%u:%s:%d:"LPX64":%s",
735 sec->ps_id, sec2gsec(sec)->gs_mech->gm_name,
736 vcred->vc_uid, vcred->vc_gid,
737 co_flags, import_to_gss_svc(imp),
738 imp->imp_connection->c_peer.nid, imp->imp_obd->obd_name);
740 CDEBUG(D_SEC, "requesting key for %s\n", desc);
742 keyring_upcall_lock(gsec_kr);
743 key = request_key(&gss_key_type, desc, coinfo);
744 keyring_upcall_unlock(gsec_kr);
746 OBD_FREE(coinfo, coinfo_size);
749 CERROR("failed request key: %ld\n", PTR_ERR(key));
752 CDEBUG(D_SEC, "obtained key %08x for %s\n", key->serial, desc);
754 /* once payload.data was pointed to a ctx, it never changes until
755 * we de-associate them; but parallel request_key() may return
756 * a key with payload.data == NULL at the same time. so we still
757 * need wirtelock of key->sem to serialize them. */
758 down_write(&key->sem);
760 if (likely(key->payload.data != NULL)) {
761 ctx = key->payload.data;
763 LASSERT(atomic_read(&ctx->cc_refcount) >= 1);
764 LASSERT(ctx2gctx_keyring(ctx)->gck_key == key);
765 LASSERT(atomic_read(&key->usage) >= 2);
767 /* simply take a ref and return. it's upper layer's
768 * responsibility to detect & replace dead ctx. */
769 atomic_inc(&ctx->cc_refcount);
771 /* pre initialization with a cli_ctx. this can't be done in
772 * key_instantiate() because we'v no enough information
774 ctx = ctx_create_kr(sec, vcred);
776 ctx_enlist_kr(ctx, is_root, 0);
777 bind_key_ctx(key, ctx);
779 ctx_start_timer_kr(ctx, KEYRING_UPCALL_TIMEOUT);
781 CDEBUG(D_SEC, "installed key %p <-> ctx %p (sec %p)\n",
784 /* we'd prefer to call key_revoke(), but we more like
785 * to revoke it within this key->sem locked period. */
786 key_revoke_locked(key);
794 if (is_root && create_new)
795 request_key_unlink(key);
800 mutex_unlock(&gsec_kr->gsk_root_uc_lock);
805 void gss_sec_release_ctx_kr(struct ptlrpc_sec *sec,
806 struct ptlrpc_cli_ctx *ctx,
809 LASSERT(atomic_read(&sec->ps_refcount) > 0);
810 LASSERT(atomic_read(&ctx->cc_refcount) == 0);
811 ctx_release_kr(ctx, sync);
815 * flush context of normal user, we must resort to keyring itself to find out
816 * contexts which belong to me.
818 * Note here we suppose only to flush _my_ context, the "uid" will
819 * be ignored in the search.
822 void flush_user_ctx_cache_kr(struct ptlrpc_sec *sec,
824 int grace, int force)
829 /* nothing to do for reverse or rootonly sec */
830 if (sec_is_reverse(sec) || sec_is_rootonly(sec))
833 construct_key_desc(desc, sizeof(desc), sec, uid);
835 /* there should be only one valid key, but we put it in the
836 * loop in case of any weird cases */
838 key = request_key(&gss_key_type, desc, NULL);
840 CDEBUG(D_SEC, "No more key found for current user\n");
844 down_write(&key->sem);
846 kill_key_locked(key);
848 /* kill_key_locked() should usually revoke the key, but we
849 * revoke it again to make sure, e.g. some case the key may
850 * not well coupled with a context. */
851 key_revoke_locked(key);
860 * flush context of root or all, we iterate through the list.
863 void flush_spec_ctx_cache_kr(struct ptlrpc_sec *sec,
865 int grace, int force)
867 struct gss_sec_keyring *gsec_kr;
868 struct hlist_head freelist = CFS_HLIST_HEAD_INIT;
869 struct hlist_node *pos, *next;
870 struct ptlrpc_cli_ctx *ctx;
873 gsec_kr = sec2gsec_keyring(sec);
875 spin_lock(&sec->ps_lock);
876 hlist_for_each_entry_safe(ctx, pos, next,
877 &gsec_kr->gsk_clist, cc_cache) {
878 LASSERT(atomic_read(&ctx->cc_refcount) > 0);
880 if (uid != -1 && uid != ctx->cc_vcred.vc_uid)
883 /* at this moment there's at least 2 base reference:
884 * key association and in-list. */
885 if (atomic_read(&ctx->cc_refcount) > 2) {
888 CWARN("flush busy ctx %p(%u->%s, extra ref %d)\n",
889 ctx, ctx->cc_vcred.vc_uid,
890 sec2target_str(ctx->cc_sec),
891 atomic_read(&ctx->cc_refcount) - 2);
894 set_bit(PTLRPC_CTX_DEAD_BIT, &ctx->cc_flags);
896 clear_bit(PTLRPC_CTX_UPTODATE_BIT, &ctx->cc_flags);
898 atomic_inc(&ctx->cc_refcount);
900 if (ctx_unlist_kr(ctx, 1)) {
901 hlist_add_head(&ctx->cc_cache, &freelist);
903 LASSERT(atomic_read(&ctx->cc_refcount) >= 2);
904 atomic_dec(&ctx->cc_refcount);
907 spin_unlock(&sec->ps_lock);
909 dispose_ctx_list_kr(&freelist);
914 int gss_sec_flush_ctx_cache_kr(struct ptlrpc_sec *sec,
916 int grace, int force)
920 CDEBUG(D_SEC, "sec %p(%d, nctx %d), uid %d, grace %d, force %d\n",
921 sec, atomic_read(&sec->ps_refcount), atomic_read(&sec->ps_nctx),
924 if (uid != -1 && uid != 0)
925 flush_user_ctx_cache_kr(sec, uid, grace, force);
927 flush_spec_ctx_cache_kr(sec, uid, grace, force);
933 void gss_sec_gc_ctx_kr(struct ptlrpc_sec *sec)
935 struct gss_sec_keyring *gsec_kr = sec2gsec_keyring(sec);
936 struct hlist_head freelist = CFS_HLIST_HEAD_INIT;
937 struct hlist_node *pos, *next;
938 struct ptlrpc_cli_ctx *ctx;
941 CWARN("running gc\n");
943 spin_lock(&sec->ps_lock);
944 hlist_for_each_entry_safe(ctx, pos, next,
945 &gsec_kr->gsk_clist, cc_cache) {
946 LASSERT(atomic_read(&ctx->cc_refcount) > 0);
948 atomic_inc(&ctx->cc_refcount);
950 if (cli_ctx_check_death(ctx) && ctx_unlist_kr(ctx, 1)) {
951 hlist_add_head(&ctx->cc_cache, &freelist);
952 CWARN("unhashed ctx %p\n", ctx);
954 LASSERT(atomic_read(&ctx->cc_refcount) >= 2);
955 atomic_dec(&ctx->cc_refcount);
958 spin_unlock(&sec->ps_lock);
960 dispose_ctx_list_kr(&freelist);
966 int gss_sec_display_kr(struct ptlrpc_sec *sec, struct seq_file *seq)
968 struct gss_sec_keyring *gsec_kr = sec2gsec_keyring(sec);
969 struct hlist_node *pos, *next;
970 struct ptlrpc_cli_ctx *ctx;
971 struct gss_cli_ctx *gctx;
972 time_t now = cfs_time_current_sec();
975 spin_lock(&sec->ps_lock);
976 hlist_for_each_entry_safe(ctx, pos, next,
977 &gsec_kr->gsk_clist, cc_cache) {
982 gctx = ctx2gctx(ctx);
983 key = ctx2gctx_keyring(ctx)->gck_key;
985 gss_cli_ctx_flags2str(ctx->cc_flags,
986 flags_str, sizeof(flags_str));
988 if (gctx->gc_mechctx)
989 lgss_display(gctx->gc_mechctx, mech, sizeof(mech));
991 snprintf(mech, sizeof(mech), "N/A");
992 mech[sizeof(mech) - 1] = '\0';
994 seq_printf(seq, "%p: uid %u, ref %d, expire %ld(%+ld), fl %s, "
995 "seq %d, win %u, key %08x(ref %d), "
996 "hdl "LPX64":"LPX64", mech: %s\n",
997 ctx, ctx->cc_vcred.vc_uid,
998 atomic_read(&ctx->cc_refcount),
1000 ctx->cc_expire ? ctx->cc_expire - now : 0,
1002 atomic_read(&gctx->gc_seq),
1004 key ? key->serial : 0,
1005 key ? atomic_read(&key->usage) : 0,
1006 gss_handle_to_u64(&gctx->gc_handle),
1007 gss_handle_to_u64(&gctx->gc_svc_handle),
1010 spin_unlock(&sec->ps_lock);
1015 /****************************************
1017 ****************************************/
1020 int gss_cli_ctx_refresh_kr(struct ptlrpc_cli_ctx *ctx)
1022 /* upcall is already on the way */
1027 int gss_cli_ctx_validate_kr(struct ptlrpc_cli_ctx *ctx)
1029 LASSERT(atomic_read(&ctx->cc_refcount) > 0);
1030 LASSERT(ctx->cc_sec);
1032 if (cli_ctx_check_death(ctx)) {
1037 if (cli_ctx_is_ready(ctx))
1043 void gss_cli_ctx_die_kr(struct ptlrpc_cli_ctx *ctx, int grace)
1045 LASSERT(atomic_read(&ctx->cc_refcount) > 0);
1046 LASSERT(ctx->cc_sec);
1048 cli_ctx_expire(ctx);
1052 /****************************************
1053 * (reverse) service *
1054 ****************************************/
1057 * reverse context could have nothing to do with keyrings. here we still keep
1058 * the version which bind to a key, for future reference.
1060 #define HAVE_REVERSE_CTX_NOKEY
1062 #ifdef HAVE_REVERSE_CTX_NOKEY
1065 int sec_install_rctx_kr(struct ptlrpc_sec *sec,
1066 struct ptlrpc_svc_ctx *svc_ctx)
1068 struct ptlrpc_cli_ctx *cli_ctx;
1069 struct vfs_cred vcred = { 0, 0 };
1075 cli_ctx = ctx_create_kr(sec, &vcred);
1076 if (cli_ctx == NULL)
1079 rc = gss_copy_rvc_cli_ctx(cli_ctx, svc_ctx);
1081 CERROR("failed copy reverse cli ctx: %d\n", rc);
1083 ctx_put_kr(cli_ctx, 1);
1087 rvs_sec_install_root_ctx_kr(sec, cli_ctx, NULL);
1089 ctx_put_kr(cli_ctx, 1);
1094 #else /* ! HAVE_REVERSE_CTX_NOKEY */
1097 int sec_install_rctx_kr(struct ptlrpc_sec *sec,
1098 struct ptlrpc_svc_ctx *svc_ctx)
1100 struct ptlrpc_cli_ctx *cli_ctx = NULL;
1102 struct vfs_cred vcred = { 0, 0 };
1110 construct_key_desc(desc, sizeof(desc), sec, 0);
1112 key = key_alloc(&gss_key_type, desc, 0, 0,
1113 KEY_POS_ALL | KEY_USR_ALL, 1);
1115 CERROR("failed to alloc key: %ld\n", PTR_ERR(key));
1116 return PTR_ERR(key);
1119 rc = key_instantiate_and_link(key, NULL, 0, NULL, NULL);
1121 CERROR("failed to instantiate key: %d\n", rc);
1125 down_write(&key->sem);
1127 LASSERT(key->payload.data == NULL);
1129 cli_ctx = ctx_create_kr(sec, &vcred);
1130 if (cli_ctx == NULL) {
1135 rc = gss_copy_rvc_cli_ctx(cli_ctx, svc_ctx);
1137 CERROR("failed copy reverse cli ctx: %d\n", rc);
1141 rvs_sec_install_root_ctx_kr(sec, cli_ctx, key);
1143 ctx_put_kr(cli_ctx, 1);
1144 up_write(&key->sem);
1153 ctx_put_kr(cli_ctx, 1);
1155 up_write(&key->sem);
1161 #endif /* HAVE_REVERSE_CTX_NOKEY */
1163 /****************************************
1165 ****************************************/
1168 int gss_svc_accept_kr(struct ptlrpc_request *req)
1170 return gss_svc_accept(&gss_policy_keyring, req);
1174 int gss_svc_install_rctx_kr(struct obd_import *imp,
1175 struct ptlrpc_svc_ctx *svc_ctx)
1177 struct ptlrpc_sec *sec;
1180 sec = sptlrpc_import_sec_ref(imp);
1183 rc = sec_install_rctx_kr(sec, svc_ctx);
1184 sptlrpc_sec_put(sec);
1189 /****************************************
1191 ****************************************/
1194 int gss_kt_instantiate(struct key *key, const void *data, size_t datalen)
1199 if (data != NULL || datalen != 0) {
1200 CERROR("invalid: data %p, len %d\n", data, datalen);
1204 if (key->payload.data != 0) {
1205 CERROR("key already have payload\n");
1209 /* link the key to session keyring, so following context negotiation
1210 * rpc fired from user space could find this key. This will be unlinked
1211 * automatically when upcall processes die.
1213 * we can't do this through keyctl from userspace, because the upcall
1214 * might be neither possessor nor owner of the key (setuid).
1216 * the session keyring is created upon upcall, and don't change all
1217 * the way until upcall finished, so rcu lock is not needed here.
1219 LASSERT(cfs_current()->signal->session_keyring);
1221 rc = key_link(cfs_current()->signal->session_keyring, key);
1223 CERROR("failed to link key %08x to keyring %08x: %d\n",
1225 cfs_current()->signal->session_keyring->serial, rc);
1229 CDEBUG(D_SEC, "key %p instantiated, ctx %p\n", key, key->payload.data);
1234 * called with key semaphore write locked. it means we can operate
1235 * on the context without fear of loosing refcount.
1238 int gss_kt_update(struct key *key, const void *data, size_t datalen)
1240 struct ptlrpc_cli_ctx *ctx = key->payload.data;
1241 struct gss_cli_ctx *gctx;
1242 rawobj_t tmpobj = RAWOBJ_EMPTY;
1246 if (data == NULL || datalen == 0) {
1247 CWARN("invalid: data %p, len %d\n", data, datalen);
1251 /* there's a race between userspace parent - child processes. if
1252 * child finish negotiation too fast and call kt_update(), the ctx
1253 * might be still NULL. but the key will finally be associate
1254 * with a context, or be revoked. if key status is fine, return
1255 * -EAGAIN to allow userspace sleep a while and call again. */
1257 CWARN("race in userspace. key %p(%x) flags %lx\n",
1258 key, key->serial, key->flags);
1260 rc = key_validate(key);
1267 LASSERT(atomic_read(&ctx->cc_refcount) > 0);
1268 LASSERT(ctx->cc_sec);
1270 ctx_clear_timer_kr(ctx);
1272 /* don't proceed if already refreshed */
1273 if (cli_ctx_is_refreshed(ctx)) {
1274 CWARN("ctx already done refresh\n");
1275 sptlrpc_cli_ctx_wakeup(ctx);
1279 sptlrpc_cli_ctx_get(ctx);
1280 gctx = ctx2gctx(ctx);
1282 rc = buffer_extract_bytes(&data, &datalen, &gctx->gc_win,
1283 sizeof(gctx->gc_win));
1285 CERROR("failed extract seq_win\n");
1289 if (gctx->gc_win == 0) {
1290 __u32 nego_rpc_err, nego_gss_err;
1292 rc = buffer_extract_bytes(&data, &datalen, &nego_rpc_err,
1293 sizeof(nego_rpc_err));
1295 CERROR("failed to extrace rpc rc\n");
1299 rc = buffer_extract_bytes(&data, &datalen, &nego_gss_err,
1300 sizeof(nego_gss_err));
1302 CERROR("failed to extrace gss rc\n");
1306 CERROR("negotiation: rpc err %d, gss err %x\n",
1307 nego_rpc_err, nego_gss_err);
1309 rc = nego_rpc_err ? nego_rpc_err : -EACCES;
1311 rc = rawobj_extract_local_alloc(&gctx->gc_handle,
1312 (__u32 **) &data, &datalen);
1314 CERROR("failed extract handle\n");
1318 rc = rawobj_extract_local(&tmpobj, (__u32 **) &data, &datalen);
1320 CERROR("failed extract mech\n");
1324 rc = lgss_import_sec_context(&tmpobj,
1325 sec2gsec(ctx->cc_sec)->gs_mech,
1327 if (rc != GSS_S_COMPLETE)
1328 CERROR("failed import context\n");
1333 /* we don't care what current status of this ctx, even someone else
1334 * is operating on the ctx at the same time. we just add up our own
1337 gss_cli_ctx_uptodate(gctx);
1339 /* this will also revoke the key. has to be done before
1340 * wakeup waiters otherwise they can find the stale key */
1341 kill_key_locked(key);
1343 cli_ctx_expire(ctx);
1345 if (rc != -ERESTART)
1346 set_bit(PTLRPC_CTX_ERROR_BIT, &ctx->cc_flags);
1349 sptlrpc_cli_ctx_wakeup(ctx);
1351 /* let user space think it's a success */
1352 sptlrpc_cli_ctx_put(ctx, 1);
1357 int gss_kt_match(const struct key *key, const void *desc)
1359 return (strcmp(key->description, (const char *) desc) == 0);
1363 void gss_kt_destroy(struct key *key)
1366 LASSERT(key->payload.data == NULL);
1367 CDEBUG(D_SEC, "destroy key %p\n", key);
1372 void gss_kt_describe(const struct key *key, struct seq_file *s)
1374 if (key->description == NULL)
1375 seq_puts(s, "[null]");
1377 seq_puts(s, key->description);
1380 static struct key_type gss_key_type =
1384 .instantiate = gss_kt_instantiate,
1385 .update = gss_kt_update,
1386 .match = gss_kt_match,
1387 .destroy = gss_kt_destroy,
1388 .describe = gss_kt_describe,
1391 /****************************************
1392 * lustre gss keyring policy *
1393 ****************************************/
1395 static struct ptlrpc_ctx_ops gss_keyring_ctxops = {
1396 .match = gss_cli_ctx_match,
1397 .refresh = gss_cli_ctx_refresh_kr,
1398 .validate = gss_cli_ctx_validate_kr,
1399 .die = gss_cli_ctx_die_kr,
1400 .sign = gss_cli_ctx_sign,
1401 .verify = gss_cli_ctx_verify,
1402 .seal = gss_cli_ctx_seal,
1403 .unseal = gss_cli_ctx_unseal,
1404 .wrap_bulk = gss_cli_ctx_wrap_bulk,
1405 .unwrap_bulk = gss_cli_ctx_unwrap_bulk,
1408 static struct ptlrpc_sec_cops gss_sec_keyring_cops = {
1409 .create_sec = gss_sec_create_kr,
1410 .destroy_sec = gss_sec_destroy_kr,
1411 .kill_sec = gss_sec_kill,
1412 .lookup_ctx = gss_sec_lookup_ctx_kr,
1413 .release_ctx = gss_sec_release_ctx_kr,
1414 .flush_ctx_cache = gss_sec_flush_ctx_cache_kr,
1415 .gc_ctx = gss_sec_gc_ctx_kr,
1416 .install_rctx = gss_sec_install_rctx,
1417 .alloc_reqbuf = gss_alloc_reqbuf,
1418 .free_reqbuf = gss_free_reqbuf,
1419 .alloc_repbuf = gss_alloc_repbuf,
1420 .free_repbuf = gss_free_repbuf,
1421 .enlarge_reqbuf = gss_enlarge_reqbuf,
1422 .display = gss_sec_display_kr,
1425 static struct ptlrpc_sec_sops gss_sec_keyring_sops = {
1426 .accept = gss_svc_accept_kr,
1427 .invalidate_ctx = gss_svc_invalidate_ctx,
1428 .alloc_rs = gss_svc_alloc_rs,
1429 .authorize = gss_svc_authorize,
1430 .free_rs = gss_svc_free_rs,
1431 .free_ctx = gss_svc_free_ctx,
1432 .unwrap_bulk = gss_svc_unwrap_bulk,
1433 .wrap_bulk = gss_svc_wrap_bulk,
1434 .install_rctx = gss_svc_install_rctx_kr,
1437 static struct ptlrpc_sec_policy gss_policy_keyring = {
1438 .sp_owner = THIS_MODULE,
1439 .sp_name = "gss.keyring",
1440 .sp_policy = SPTLRPC_POLICY_GSS,
1441 .sp_cops = &gss_sec_keyring_cops,
1442 .sp_sops = &gss_sec_keyring_sops,
1446 int __init gss_init_keyring(void)
1450 rc = register_key_type(&gss_key_type);
1452 CERROR("failed to register keyring type: %d\n", rc);
1456 rc = sptlrpc_register_policy(&gss_policy_keyring);
1458 unregister_key_type(&gss_key_type);
1465 void __exit gss_exit_keyring(void)
1467 unregister_key_type(&gss_key_type);
1468 sptlrpc_unregister_policy(&gss_policy_keyring);