4 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
6 * This program is free software; you can redistribute it and/or modify
7 * it under the terms of the GNU General Public License version 2 only,
8 * as published by the Free Software Foundation.
10 * This program is distributed in the hope that it will be useful, but
11 * WITHOUT ANY WARRANTY; without even the implied warranty of
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
13 * General Public License version 2 for more details (a copy is included
14 * in the LICENSE file that accompanied this code).
16 * You should have received a copy of the GNU General Public License
17 * version 2 along with this program; If not, see
18 * http://www.gnu.org/licenses/gpl-2.0.html
23 * Copyright (c) 2007, 2010, Oracle and/or its affiliates. All rights reserved.
24 * Use is subject to license terms.
26 * Copyright (c) 2012, 2014, Intel Corporation.
29 * This file is part of Lustre, http://www.lustre.org/
31 * lustre/ptlrpc/gss/gss_keyring.c
33 * Author: Eric Mei <ericm@clusterfs.com>
36 #define DEBUG_SUBSYSTEM S_SEC
37 #include <linux/init.h>
38 #include <linux/module.h>
39 #include <linux/slab.h>
40 #include <linux/dcache.h>
42 #include <linux/crypto.h>
43 #include <linux/key.h>
44 #include <linux/keyctl.h>
45 #include <linux/key-type.h>
46 #include <linux/mutex.h>
47 #include <asm/atomic.h>
49 #include <libcfs/linux/linux-list.h>
51 #include <obd_class.h>
52 #include <obd_support.h>
53 #include <uapi/linux/lustre/lustre_idl.h>
54 #include <lustre_sec.h>
55 #include <lustre_net.h>
56 #include <lustre_import.h>
59 #include "gss_internal.h"
62 #ifdef HAVE_GET_REQUEST_KEY_AUTH
63 #include <keys/request_key_auth-type.h>
66 static struct ptlrpc_sec_policy gss_policy_keyring;
67 static struct ptlrpc_ctx_ops gss_keyring_ctxops;
68 static struct key_type gss_key_type;
70 static int sec_install_rctx_kr(struct ptlrpc_sec *sec,
71 struct ptlrpc_svc_ctx *svc_ctx);
72 static void request_key_unlink(struct key *key);
75 * the timeout is only for the case that upcall child process die abnormally.
76 * in any other cases it should finally update kernel key.
78 * FIXME we'd better to incorporate the client & server side upcall timeouts
79 * into the framework of Adaptive Timeouts, but we need to figure out how to
80 * make sure that kernel knows the upcall processes is in-progress or died
83 #define KEYRING_UPCALL_TIMEOUT (obd_timeout + obd_timeout)
85 /* Check caller's namespace in gss_keyring upcall */
86 unsigned int gss_check_upcall_ns = 1;
88 /****************************************
90 ****************************************/
92 static inline void keyring_upcall_lock(struct gss_sec_keyring *gsec_kr)
94 #ifdef HAVE_KEYRING_UPCALL_SERIALIZED
95 mutex_lock(&gsec_kr->gsk_uc_lock);
99 static inline void keyring_upcall_unlock(struct gss_sec_keyring *gsec_kr)
101 #ifdef HAVE_KEYRING_UPCALL_SERIALIZED
102 mutex_unlock(&gsec_kr->gsk_uc_lock);
106 static inline void key_revoke_locked(struct key *key)
108 set_bit(KEY_FLAG_REVOKED, &key->flags);
111 static void ctx_upcall_timeout_kr(cfs_timer_cb_arg_t data)
113 struct gss_cli_ctx_keyring *gctx_kr = cfs_from_timer(gctx_kr,
115 struct ptlrpc_cli_ctx *ctx = &(gctx_kr->gck_base.gc_base);
116 struct key *key = gctx_kr->gck_key;
118 CWARN("ctx %p, key %p\n", ctx, key);
123 key_revoke_locked(key);
126 static void ctx_start_timer_kr(struct ptlrpc_cli_ctx *ctx, time64_t timeout)
128 struct gss_cli_ctx_keyring *gctx_kr = ctx2gctx_keyring(ctx);
129 struct timer_list *timer = &gctx_kr->gck_timer;
133 CDEBUG(D_SEC, "ctx %p: start timer %llds\n", ctx, timeout);
135 cfs_timer_setup(timer, ctx_upcall_timeout_kr,
136 (unsigned long)gctx_kr, 0);
137 timer->expires = cfs_time_seconds(timeout) + jiffies;
142 * caller should make sure no race with other threads
145 void ctx_clear_timer_kr(struct ptlrpc_cli_ctx *ctx)
147 struct gss_cli_ctx_keyring *gctx_kr = ctx2gctx_keyring(ctx);
148 struct timer_list *timer = &gctx_kr->gck_timer;
150 CDEBUG(D_SEC, "ctx %p, key %p\n", ctx, gctx_kr->gck_key);
152 del_singleshot_timer_sync(timer);
156 struct ptlrpc_cli_ctx *ctx_create_kr(struct ptlrpc_sec *sec,
157 struct vfs_cred *vcred)
159 struct ptlrpc_cli_ctx *ctx;
160 struct gss_cli_ctx_keyring *gctx_kr;
162 OBD_ALLOC_PTR(gctx_kr);
166 cfs_timer_setup(&gctx_kr->gck_timer, NULL, 0, 0);
168 ctx = &gctx_kr->gck_base.gc_base;
170 if (gss_cli_ctx_init_common(sec, ctx, &gss_keyring_ctxops, vcred)) {
171 OBD_FREE_PTR(gctx_kr);
175 ctx->cc_expire = ktime_get_real_seconds() + KEYRING_UPCALL_TIMEOUT;
176 clear_bit(PTLRPC_CTX_NEW_BIT, &ctx->cc_flags);
177 atomic_inc(&ctx->cc_refcount); /* for the caller */
182 static void ctx_destroy_kr(struct ptlrpc_cli_ctx *ctx)
184 struct ptlrpc_sec *sec = ctx->cc_sec;
185 struct gss_cli_ctx_keyring *gctx_kr = ctx2gctx_keyring(ctx);
187 CDEBUG(D_SEC, "destroying ctx %p\n", ctx);
189 /* at this time the association with key has been broken. */
191 LASSERT(atomic_read(&sec->ps_refcount) > 0);
192 LASSERT(atomic_read(&sec->ps_nctx) > 0);
193 LASSERT(test_bit(PTLRPC_CTX_CACHED_BIT, &ctx->cc_flags) == 0);
194 LASSERT(gctx_kr->gck_key == NULL);
196 ctx_clear_timer_kr(ctx);
198 if (gss_cli_ctx_fini_common(sec, ctx))
201 OBD_FREE_PTR(gctx_kr);
203 atomic_dec(&sec->ps_nctx);
204 sptlrpc_sec_put(sec);
207 static void ctx_release_kr(struct ptlrpc_cli_ctx *ctx, int sync)
212 atomic_inc(&ctx->cc_refcount);
213 sptlrpc_gc_add_ctx(ctx);
217 static void ctx_put_kr(struct ptlrpc_cli_ctx *ctx, int sync)
219 LASSERT(atomic_read(&ctx->cc_refcount) > 0);
221 if (atomic_dec_and_test(&ctx->cc_refcount))
222 ctx_release_kr(ctx, sync);
226 * key <-> ctx association and rules:
227 * - ctx might not bind with any key
228 * - key/ctx binding is protected by key semaphore (if the key present)
229 * - key and ctx each take a reference of the other
230 * - ctx enlist/unlist is protected by ctx spinlock
231 * - never enlist a ctx after it's been unlisted
232 * - whoever do enlist should also do bind, lock key before enlist:
233 * - lock key -> lock ctx -> enlist -> unlock ctx -> bind -> unlock key
234 * - whoever do unlist should also do unbind:
235 * - lock key -> lock ctx -> unlist -> unlock ctx -> unbind -> unlock key
236 * - lock ctx -> unlist -> unlock ctx -> lock key -> unbind -> unlock key
239 static inline void spin_lock_if(spinlock_t *lock, int condition)
245 static inline void spin_unlock_if(spinlock_t *lock, int condition)
251 static void ctx_enlist_kr(struct ptlrpc_cli_ctx *ctx, int is_root, int locked)
253 struct ptlrpc_sec *sec = ctx->cc_sec;
254 struct gss_sec_keyring *gsec_kr = sec2gsec_keyring(sec);
256 LASSERT(!test_bit(PTLRPC_CTX_CACHED_BIT, &ctx->cc_flags));
257 LASSERT(atomic_read(&ctx->cc_refcount) > 0);
259 spin_lock_if(&sec->ps_lock, !locked);
261 atomic_inc(&ctx->cc_refcount);
262 set_bit(PTLRPC_CTX_CACHED_BIT, &ctx->cc_flags);
263 hlist_add_head(&ctx->cc_cache, &gsec_kr->gsk_clist);
265 gsec_kr->gsk_root_ctx = ctx;
267 spin_unlock_if(&sec->ps_lock, !locked);
271 * Note after this get called, caller should not access ctx again because
272 * it might have been freed, unless caller hold at least one refcount of
275 * return non-zero if we indeed unlist this ctx.
277 static int ctx_unlist_kr(struct ptlrpc_cli_ctx *ctx, int locked)
279 struct ptlrpc_sec *sec = ctx->cc_sec;
280 struct gss_sec_keyring *gsec_kr = sec2gsec_keyring(sec);
282 /* if hashed bit has gone, leave the job to somebody who is doing it */
283 if (test_and_clear_bit(PTLRPC_CTX_CACHED_BIT, &ctx->cc_flags) == 0)
286 /* drop ref inside spin lock to prevent race with other operations */
287 spin_lock_if(&sec->ps_lock, !locked);
289 if (gsec_kr->gsk_root_ctx == ctx)
290 gsec_kr->gsk_root_ctx = NULL;
291 hlist_del_init(&ctx->cc_cache);
292 atomic_dec(&ctx->cc_refcount);
294 spin_unlock_if(&sec->ps_lock, !locked);
300 * Get specific payload. Newer kernels support 4 slots.
303 key_get_payload(struct key *key, unsigned int index)
305 void *key_ptr = NULL;
307 #ifdef HAVE_KEY_PAYLOAD_DATA_ARRAY
308 key_ptr = key->payload.data[index];
311 key_ptr = key->payload.data;
317 * Set specific payload. Newer kernels support 4 slots.
319 static int key_set_payload(struct key *key, unsigned int index,
320 struct ptlrpc_cli_ctx *ctx)
324 #ifdef HAVE_KEY_PAYLOAD_DATA_ARRAY
326 key->payload.data[index] = ctx;
329 key->payload.data = ctx;
337 * bind a key with a ctx together.
338 * caller must hold write lock of the key, as well as ref on key & ctx.
340 static void bind_key_ctx(struct key *key, struct ptlrpc_cli_ctx *ctx)
342 LASSERT(atomic_read(&ctx->cc_refcount) > 0);
343 LASSERT(ll_read_key_usage(key) > 0);
344 LASSERT(ctx2gctx_keyring(ctx)->gck_key == NULL);
345 LASSERT(!key_get_payload(key, 0));
347 /* at this time context may or may not in list. */
349 atomic_inc(&ctx->cc_refcount);
350 ctx2gctx_keyring(ctx)->gck_key = key;
351 LASSERT(!key_set_payload(key, 0, ctx));
355 * unbind a key and a ctx.
356 * caller must hold write lock, as well as a ref of the key.
358 static void unbind_key_ctx(struct key *key, struct ptlrpc_cli_ctx *ctx)
360 LASSERT(key_get_payload(key, 0) == ctx);
361 LASSERT(test_bit(PTLRPC_CTX_CACHED_BIT, &ctx->cc_flags) == 0);
363 /* must revoke the key, or others may treat it as newly created */
364 key_revoke_locked(key);
366 key_set_payload(key, 0, NULL);
367 ctx2gctx_keyring(ctx)->gck_key = NULL;
369 /* once ctx get split from key, the timer is meaningless */
370 ctx_clear_timer_kr(ctx);
377 * given a ctx, unbind with its coupled key, if any.
378 * unbind could only be called once, so we don't worry the key be released
381 static void unbind_ctx_kr(struct ptlrpc_cli_ctx *ctx)
383 struct key *key = ctx2gctx_keyring(ctx)->gck_key;
386 LASSERT(key_get_payload(key, 0) == ctx);
389 down_write(&key->sem);
390 unbind_key_ctx(key, ctx);
393 request_key_unlink(key);
398 * given a key, unbind with its coupled ctx, if any.
399 * caller must hold write lock, as well as a ref of the key.
401 static void unbind_key_locked(struct key *key)
403 struct ptlrpc_cli_ctx *ctx = key_get_payload(key, 0);
406 unbind_key_ctx(key, ctx);
410 * unlist a ctx, and unbind from coupled key
412 static void kill_ctx_kr(struct ptlrpc_cli_ctx *ctx)
414 if (ctx_unlist_kr(ctx, 0))
419 * given a key, unlist and unbind with the coupled ctx (if any).
420 * caller must hold write lock, as well as a ref of the key.
422 static void kill_key_locked(struct key *key)
424 struct ptlrpc_cli_ctx *ctx = key_get_payload(key, 0);
426 if (ctx && ctx_unlist_kr(ctx, 0))
427 unbind_key_locked(key);
431 * caller should hold one ref on contexts in freelist.
433 static void dispose_ctx_list_kr(struct hlist_head *freelist)
435 struct hlist_node *next;
436 struct ptlrpc_cli_ctx *ctx;
437 struct gss_cli_ctx *gctx;
439 hlist_for_each_entry_safe(ctx, next, freelist, cc_cache) {
440 hlist_del_init(&ctx->cc_cache);
442 /* reverse ctx: update current seq to buddy svcctx if exist.
443 * ideally this should be done at gss_cli_ctx_finalize(), but
444 * the ctx destroy could be delayed by:
445 * 1) ctx still has reference;
446 * 2) ctx destroy is asynchronous;
447 * and reverse import call inval_all_ctx() require this be done
448 * _immediately_ otherwise newly created reverse ctx might copy
449 * the very old sequence number from svcctx. */
450 gctx = ctx2gctx(ctx);
451 if (!rawobj_empty(&gctx->gc_svc_handle) &&
452 sec_is_reverse(gctx->gc_base.cc_sec)) {
453 gss_svc_upcall_update_sequence(&gctx->gc_svc_handle,
454 (__u32) atomic_read(&gctx->gc_seq));
457 /* we need to wakeup waiting reqs here. the context might
458 * be forced released before upcall finished, then the
459 * late-arrived downcall can't find the ctx even. */
460 sptlrpc_cli_ctx_wakeup(ctx);
468 * lookup a root context directly in a sec, return root ctx with a
469 * reference taken or NULL.
472 struct ptlrpc_cli_ctx * sec_lookup_root_ctx_kr(struct ptlrpc_sec *sec)
474 struct gss_sec_keyring *gsec_kr = sec2gsec_keyring(sec);
475 struct ptlrpc_cli_ctx *ctx = NULL;
477 spin_lock(&sec->ps_lock);
479 ctx = gsec_kr->gsk_root_ctx;
481 if (ctx == NULL && unlikely(sec_is_reverse(sec))) {
482 struct ptlrpc_cli_ctx *tmp;
484 /* reverse ctx, search root ctx in list, choose the one
485 * with shortest expire time, which is most possibly have
486 * an established peer ctx at client side. */
487 hlist_for_each_entry(tmp, &gsec_kr->gsk_clist, cc_cache) {
488 if (ctx == NULL || ctx->cc_expire == 0 ||
489 ctx->cc_expire > tmp->cc_expire) {
491 /* promote to be root_ctx */
492 gsec_kr->gsk_root_ctx = ctx;
498 LASSERT(atomic_read(&ctx->cc_refcount) > 0);
499 LASSERT(!hlist_empty(&gsec_kr->gsk_clist));
500 atomic_inc(&ctx->cc_refcount);
503 spin_unlock(&sec->ps_lock);
508 #define RVS_CTX_EXPIRE_NICE (10)
511 void rvs_sec_install_root_ctx_kr(struct ptlrpc_sec *sec,
512 struct ptlrpc_cli_ctx *new_ctx,
515 struct gss_sec_keyring *gsec_kr = sec2gsec_keyring(sec);
516 struct ptlrpc_cli_ctx *ctx;
520 LASSERT(sec_is_reverse(sec));
522 spin_lock(&sec->ps_lock);
524 now = ktime_get_real_seconds();
526 /* set all existing ctxs short expiry */
527 hlist_for_each_entry(ctx, &gsec_kr->gsk_clist, cc_cache) {
528 if (ctx->cc_expire > now + RVS_CTX_EXPIRE_NICE) {
529 ctx->cc_early_expire = 1;
530 ctx->cc_expire = now + RVS_CTX_EXPIRE_NICE;
534 /* if there's root_ctx there, instead obsolete the current
535 * immediately, we leave it continue operating for a little while.
536 * hopefully when the first backward rpc with newest ctx send out,
537 * the client side already have the peer ctx well established. */
538 ctx_enlist_kr(new_ctx, gsec_kr->gsk_root_ctx ? 0 : 1, 1);
541 bind_key_ctx(key, new_ctx);
543 spin_unlock(&sec->ps_lock);
546 static void construct_key_desc(void *buf, int bufsize,
547 struct ptlrpc_sec *sec, uid_t uid)
549 snprintf(buf, bufsize, "%d@%x", uid, sec->ps_id);
550 ((char *)buf)[bufsize - 1] = '\0';
553 /****************************************
555 ****************************************/
558 struct ptlrpc_sec * gss_sec_create_kr(struct obd_import *imp,
559 struct ptlrpc_svc_ctx *svcctx,
560 struct sptlrpc_flavor *sf)
562 struct gss_sec_keyring *gsec_kr;
565 OBD_ALLOC(gsec_kr, sizeof(*gsec_kr));
569 INIT_HLIST_HEAD(&gsec_kr->gsk_clist);
570 gsec_kr->gsk_root_ctx = NULL;
571 mutex_init(&gsec_kr->gsk_root_uc_lock);
572 #ifdef HAVE_KEYRING_UPCALL_SERIALIZED
573 mutex_init(&gsec_kr->gsk_uc_lock);
576 if (gss_sec_create_common(&gsec_kr->gsk_base, &gss_policy_keyring,
580 if (svcctx != NULL &&
581 sec_install_rctx_kr(&gsec_kr->gsk_base.gs_base, svcctx)) {
582 gss_sec_destroy_common(&gsec_kr->gsk_base);
586 RETURN(&gsec_kr->gsk_base.gs_base);
589 OBD_FREE(gsec_kr, sizeof(*gsec_kr));
594 void gss_sec_destroy_kr(struct ptlrpc_sec *sec)
596 struct gss_sec *gsec = sec2gsec(sec);
597 struct gss_sec_keyring *gsec_kr = sec2gsec_keyring(sec);
599 CDEBUG(D_SEC, "destroy %s@%p\n", sec->ps_policy->sp_name, sec);
601 LASSERT(hlist_empty(&gsec_kr->gsk_clist));
602 LASSERT(gsec_kr->gsk_root_ctx == NULL);
604 gss_sec_destroy_common(gsec);
606 OBD_FREE(gsec_kr, sizeof(*gsec_kr));
609 static inline int user_is_root(struct ptlrpc_sec *sec, struct vfs_cred *vcred)
611 /* except the ROOTONLY flag, treat it as root user only if real uid
612 * is 0, euid/fsuid being 0 are handled as setuid scenarios */
613 if (sec_is_rootonly(sec) || (vcred->vc_uid == 0))
620 * kernel 5.3: commit 0f44e4d976f96c6439da0d6717238efa4b91196e
621 * keys: Move the user and user-session keyrings to the user_namespace
623 * When lookup_user_key is available use the kernel API rather than directly
624 * accessing the uid_keyring and session_keyring via the current process
627 #ifdef HAVE_LOOKUP_USER_KEY
629 /* from Linux security/keys/internal.h: */
630 #ifndef KEY_LOOKUP_FOR_UNLINK
631 #define KEY_LOOKUP_FOR_UNLINK 0x04
634 static struct key *_user_key(key_serial_t id)
639 ref = lookup_user_key(id, KEY_LOOKUP_FOR_UNLINK, 0);
642 return key_ref_to_ptr(ref);
645 static inline struct key *get_user_session_keyring(const struct cred *cred)
647 return _user_key(KEY_SPEC_USER_SESSION_KEYRING);
650 static inline struct key *get_user_keyring(const struct cred *cred)
652 return _user_key(KEY_SPEC_USER_KEYRING);
655 static inline struct key *get_user_session_keyring(const struct cred *cred)
657 return key_get(cred->user->session_keyring);
660 static inline struct key *get_user_keyring(const struct cred *cred)
662 return key_get(cred->user->uid_keyring);
667 * unlink request key from it's ring, which is linked during request_key().
668 * sadly, we have to 'guess' which keyring it's linked to.
670 * FIXME this code is fragile, it depends on how request_key() is implemented.
672 static void request_key_unlink(struct key *key)
674 const struct cred *cred = current_cred();
675 struct key *ring = NULL;
677 switch (cred->jit_keyring) {
678 case KEY_REQKEY_DEFL_DEFAULT:
679 case KEY_REQKEY_DEFL_REQUESTOR_KEYRING:
680 #ifdef HAVE_GET_REQUEST_KEY_AUTH
681 if (cred->request_key_auth) {
682 struct request_key_auth *rka;
683 struct key *authkey = cred->request_key_auth;
685 down_read(&authkey->sem);
686 rka = get_request_key_auth(authkey);
687 if (!test_bit(KEY_FLAG_REVOKED, &authkey->flags))
688 ring = key_get(rka->dest_keyring);
689 up_read(&authkey->sem);
695 case KEY_REQKEY_DEFL_THREAD_KEYRING:
696 ring = key_get(cred->thread_keyring);
700 case KEY_REQKEY_DEFL_PROCESS_KEYRING:
701 ring = key_get(cred->process_keyring);
705 case KEY_REQKEY_DEFL_SESSION_KEYRING:
707 ring = key_get(rcu_dereference(cred->session_keyring));
712 case KEY_REQKEY_DEFL_USER_SESSION_KEYRING:
713 ring = get_user_session_keyring(cred);
715 case KEY_REQKEY_DEFL_USER_KEYRING:
716 ring = get_user_keyring(cred);
718 case KEY_REQKEY_DEFL_GROUP_KEYRING:
724 key_unlink(ring, key);
729 struct ptlrpc_cli_ctx * gss_sec_lookup_ctx_kr(struct ptlrpc_sec *sec,
730 struct vfs_cred *vcred,
731 int create, int remove_dead)
733 struct obd_import *imp = sec->ps_import;
734 struct gss_sec_keyring *gsec_kr = sec2gsec_keyring(sec);
735 struct ptlrpc_cli_ctx *ctx = NULL;
736 unsigned int is_root = 0, create_new = 0;
741 const char *sec_part_flags = "";
746 LASSERT(imp != NULL);
748 is_root = user_is_root(sec, vcred);
750 /* a little bit optimization for root context */
752 ctx = sec_lookup_root_ctx_kr(sec);
754 * Only lookup directly for REVERSE sec, which should
757 if (ctx || sec_is_reverse(sec))
761 LASSERT(create != 0);
763 /* for root context, obtain lock and check again, this time hold
764 * the root upcall lock, make sure nobody else populated new root
765 * context after last check.
768 mutex_lock(&gsec_kr->gsk_root_uc_lock);
770 ctx = sec_lookup_root_ctx_kr(sec);
774 /* update reverse handle for root user */
775 sec2gsec(sec)->gs_rvs_hdl = gss_get_next_ctx_index();
777 switch (sec->ps_part) {
779 sec_part_flags = "m";
782 sec_part_flags = "o";
785 sec_part_flags = "rmo";
788 sec_part_flags = "r";
795 switch (SPTLRPC_FLVR_SVC(sec->ps_flvr.sf_rpc)) {
796 case SPTLRPC_SVC_NULL:
799 case SPTLRPC_SVC_AUTH:
802 case SPTLRPC_SVC_INTG:
805 case SPTLRPC_SVC_PRIV:
813 /* in case of setuid, key will be constructed as owner of fsuid/fsgid,
814 * but we do authentication based on real uid/gid. the key permission
815 * bits will be exactly as POS_ALL, so only processes who subscribed
816 * this key could have the access, although the quota might be counted
817 * on others (fsuid/fsgid).
819 * keyring will use fsuid/fsgid as upcall parameters, so we have to
820 * encode real uid/gid into callout info.
823 /* But first we need to make sure the obd type is supported */
824 if (strcmp(imp->imp_obd->obd_type->typ_name, LUSTRE_MDC_NAME) &&
825 strcmp(imp->imp_obd->obd_type->typ_name, LUSTRE_OSC_NAME) &&
826 strcmp(imp->imp_obd->obd_type->typ_name, LUSTRE_MGC_NAME) &&
827 strcmp(imp->imp_obd->obd_type->typ_name, LUSTRE_LWP_NAME) &&
828 strcmp(imp->imp_obd->obd_type->typ_name, LUSTRE_OSP_NAME)) {
829 CERROR("obd %s is not a supported device\n",
830 imp->imp_obd->obd_name);
831 GOTO(out, ctx = NULL);
834 construct_key_desc(desc, sizeof(desc), sec, vcred->vc_uid);
836 /* callout info format:
837 * secid:mech:uid:gid:sec_flags:svc_flag:svc_type:peer_nid:target_uuid:
840 coinfo_size = sizeof(struct obd_uuid) + MAX_OBD_NAME + 64;
841 OBD_ALLOC(coinfo, coinfo_size);
845 /* Last callout parameter is pid of process whose namespace will be used
846 * for credentials' retrieval.
848 if (gss_check_upcall_ns) {
849 /* For user's credentials (in which case sec_part_flags is
850 * empty), use current PID instead of import's reference
851 * PID to get reference namespace.
853 if (sec_part_flags[0] == '\0')
854 caller_pid = current->pid;
856 caller_pid = imp->imp_sec_refpid;
858 /* Do not switch namespace in gss keyring upcall. */
861 snprintf(coinfo, coinfo_size, "%d:%s:%u:%u:%s:%c:%d:%#llx:%s:%#llx:%d",
862 sec->ps_id, sec2gsec(sec)->gs_mech->gm_name,
863 vcred->vc_uid, vcred->vc_gid,
864 sec_part_flags, svc_flag, import_to_gss_svc(imp),
865 imp->imp_connection->c_peer.nid, imp->imp_obd->obd_name,
866 LNetPrimaryNID(imp->imp_connection->c_self), caller_pid);
868 CDEBUG(D_SEC, "requesting key for %s\n", desc);
870 keyring_upcall_lock(gsec_kr);
871 key = request_key(&gss_key_type, desc, coinfo);
872 keyring_upcall_unlock(gsec_kr);
874 OBD_FREE(coinfo, coinfo_size);
877 CERROR("failed request key: %ld\n", PTR_ERR(key));
880 CDEBUG(D_SEC, "obtained key %08x for %s\n", key->serial, desc);
882 /* once payload.data was pointed to a ctx, it never changes until
883 * we de-associate them; but parallel request_key() may return
884 * a key with payload.data == NULL at the same time. so we still
885 * need wirtelock of key->sem to serialize them.
887 down_write(&key->sem);
889 ctx = key_get_payload(key, 0);
891 LASSERT(atomic_read(&ctx->cc_refcount) >= 1);
892 LASSERT(ctx2gctx_keyring(ctx)->gck_key == key);
893 LASSERT(ll_read_key_usage(key) >= 2);
895 /* simply take a ref and return. it's upper layer's
896 * responsibility to detect & replace dead ctx.
898 atomic_inc(&ctx->cc_refcount);
900 /* pre initialization with a cli_ctx. this can't be done in
901 * key_instantiate() because we'v no enough information
904 ctx = ctx_create_kr(sec, vcred);
906 ctx_enlist_kr(ctx, is_root, 0);
907 bind_key_ctx(key, ctx);
909 ctx_start_timer_kr(ctx, KEYRING_UPCALL_TIMEOUT);
911 CDEBUG(D_SEC, "installed key %p <-> ctx %p (sec %p)\n",
914 /* we'd prefer to call key_revoke(), but we more like
915 * to revoke it within this key->sem locked period.
917 key_revoke_locked(key);
925 if (is_root && create_new)
926 request_key_unlink(key);
931 mutex_unlock(&gsec_kr->gsk_root_uc_lock);
936 void gss_sec_release_ctx_kr(struct ptlrpc_sec *sec,
937 struct ptlrpc_cli_ctx *ctx,
940 LASSERT(atomic_read(&sec->ps_refcount) > 0);
941 LASSERT(atomic_read(&ctx->cc_refcount) == 0);
942 ctx_release_kr(ctx, sync);
946 * flush context of normal user, we must resort to keyring itself to find out
947 * contexts which belong to me.
949 * Note here we suppose only to flush _my_ context, the "uid" will
950 * be ignored in the search.
953 void flush_user_ctx_cache_kr(struct ptlrpc_sec *sec,
955 int grace, int force)
960 /* nothing to do for reverse or rootonly sec */
961 if (sec_is_reverse(sec) || sec_is_rootonly(sec))
964 construct_key_desc(desc, sizeof(desc), sec, uid);
966 /* there should be only one valid key, but we put it in the
967 * loop in case of any weird cases */
969 key = request_key(&gss_key_type, desc, NULL);
971 CDEBUG(D_SEC, "No more key found for current user\n");
975 down_write(&key->sem);
977 kill_key_locked(key);
979 /* kill_key_locked() should usually revoke the key, but we
980 * revoke it again to make sure, e.g. some case the key may
981 * not well coupled with a context. */
982 key_revoke_locked(key);
986 request_key_unlink(key);
993 * flush context of root or all, we iterate through the list.
996 void flush_spec_ctx_cache_kr(struct ptlrpc_sec *sec, uid_t uid, int grace,
999 struct gss_sec_keyring *gsec_kr;
1000 struct hlist_head freelist = HLIST_HEAD_INIT;
1001 struct hlist_node *next;
1002 struct ptlrpc_cli_ctx *ctx;
1005 gsec_kr = sec2gsec_keyring(sec);
1007 spin_lock(&sec->ps_lock);
1008 hlist_for_each_entry_safe(ctx, next, &gsec_kr->gsk_clist,
1010 LASSERT(atomic_read(&ctx->cc_refcount) > 0);
1012 if (uid != -1 && uid != ctx->cc_vcred.vc_uid)
1015 /* at this moment there's at least 2 base reference:
1016 * key association and in-list. */
1017 if (atomic_read(&ctx->cc_refcount) > 2) {
1020 CWARN("flush busy ctx %p(%u->%s, extra ref %d)\n",
1021 ctx, ctx->cc_vcred.vc_uid,
1022 sec2target_str(ctx->cc_sec),
1023 atomic_read(&ctx->cc_refcount) - 2);
1026 set_bit(PTLRPC_CTX_DEAD_BIT, &ctx->cc_flags);
1028 clear_bit(PTLRPC_CTX_UPTODATE_BIT, &ctx->cc_flags);
1030 atomic_inc(&ctx->cc_refcount);
1032 if (ctx_unlist_kr(ctx, 1)) {
1033 hlist_add_head(&ctx->cc_cache, &freelist);
1035 LASSERT(atomic_read(&ctx->cc_refcount) >= 2);
1036 atomic_dec(&ctx->cc_refcount);
1039 spin_unlock(&sec->ps_lock);
1041 dispose_ctx_list_kr(&freelist);
1046 int gss_sec_flush_ctx_cache_kr(struct ptlrpc_sec *sec,
1047 uid_t uid, int grace, int force)
1051 CDEBUG(D_SEC, "sec %p(%d, nctx %d), uid %d, grace %d, force %d\n",
1052 sec, atomic_read(&sec->ps_refcount),
1053 atomic_read(&sec->ps_nctx),
1056 if (uid != -1 && uid != 0)
1057 flush_user_ctx_cache_kr(sec, uid, grace, force);
1059 flush_spec_ctx_cache_kr(sec, uid, grace, force);
1065 void gss_sec_gc_ctx_kr(struct ptlrpc_sec *sec)
1067 struct gss_sec_keyring *gsec_kr = sec2gsec_keyring(sec);
1068 struct hlist_head freelist = HLIST_HEAD_INIT;
1069 struct hlist_node *next;
1070 struct ptlrpc_cli_ctx *ctx;
1073 CWARN("running gc\n");
1075 spin_lock(&sec->ps_lock);
1076 hlist_for_each_entry_safe(ctx, next, &gsec_kr->gsk_clist,
1078 LASSERT(atomic_read(&ctx->cc_refcount) > 0);
1080 atomic_inc(&ctx->cc_refcount);
1082 if (cli_ctx_check_death(ctx) && ctx_unlist_kr(ctx, 1)) {
1083 hlist_add_head(&ctx->cc_cache, &freelist);
1084 CWARN("unhashed ctx %p\n", ctx);
1086 LASSERT(atomic_read(&ctx->cc_refcount) >= 2);
1087 atomic_dec(&ctx->cc_refcount);
1090 spin_unlock(&sec->ps_lock);
1092 dispose_ctx_list_kr(&freelist);
1097 int gss_sec_display_kr(struct ptlrpc_sec *sec, struct seq_file *seq)
1099 struct gss_sec_keyring *gsec_kr = sec2gsec_keyring(sec);
1100 struct hlist_node *next;
1101 struct ptlrpc_cli_ctx *ctx;
1102 struct gss_cli_ctx *gctx;
1103 time64_t now = ktime_get_real_seconds();
1106 spin_lock(&sec->ps_lock);
1107 hlist_for_each_entry_safe(ctx, next, &gsec_kr->gsk_clist,
1113 gctx = ctx2gctx(ctx);
1114 key = ctx2gctx_keyring(ctx)->gck_key;
1116 gss_cli_ctx_flags2str(ctx->cc_flags,
1117 flags_str, sizeof(flags_str));
1119 if (gctx->gc_mechctx)
1120 lgss_display(gctx->gc_mechctx, mech, sizeof(mech));
1122 snprintf(mech, sizeof(mech), "N/A");
1123 mech[sizeof(mech) - 1] = '\0';
1126 "%p: uid %u, ref %d, expire %lld(%+lld), fl %s, seq %d, win %u, key %08x(ref %d), hdl %#llx:%#llx, mech: %s\n",
1127 ctx, ctx->cc_vcred.vc_uid,
1128 atomic_read(&ctx->cc_refcount),
1130 ctx->cc_expire ? ctx->cc_expire - now : 0,
1132 atomic_read(&gctx->gc_seq),
1134 key ? key->serial : 0,
1135 key ? ll_read_key_usage(key) : 0,
1136 gss_handle_to_u64(&gctx->gc_handle),
1137 gss_handle_to_u64(&gctx->gc_svc_handle),
1140 spin_unlock(&sec->ps_lock);
1145 /****************************************
1147 ****************************************/
1150 int gss_cli_ctx_refresh_kr(struct ptlrpc_cli_ctx *ctx)
1152 /* upcall is already on the way */
1153 struct gss_cli_ctx *gctx = ctx ? ctx2gctx(ctx) : NULL;
1155 /* record latest sequence number in buddy svcctx */
1156 if (gctx && !rawobj_empty(&gctx->gc_svc_handle) &&
1157 sec_is_reverse(gctx->gc_base.cc_sec)) {
1158 return gss_svc_upcall_update_sequence(&gctx->gc_svc_handle,
1159 (__u32)atomic_read(&gctx->gc_seq));
1165 int gss_cli_ctx_validate_kr(struct ptlrpc_cli_ctx *ctx)
1167 LASSERT(atomic_read(&ctx->cc_refcount) > 0);
1168 LASSERT(ctx->cc_sec);
1170 if (cli_ctx_check_death(ctx)) {
1175 if (cli_ctx_is_ready(ctx))
1181 void gss_cli_ctx_die_kr(struct ptlrpc_cli_ctx *ctx, int grace)
1183 LASSERT(atomic_read(&ctx->cc_refcount) > 0);
1184 LASSERT(ctx->cc_sec);
1186 cli_ctx_expire(ctx);
1190 /****************************************
1191 * (reverse) service *
1192 ****************************************/
1195 * reverse context could have nothing to do with keyrings. here we still keep
1196 * the version which bind to a key, for future reference.
1198 #define HAVE_REVERSE_CTX_NOKEY
1200 #ifdef HAVE_REVERSE_CTX_NOKEY
1203 int sec_install_rctx_kr(struct ptlrpc_sec *sec,
1204 struct ptlrpc_svc_ctx *svc_ctx)
1206 struct ptlrpc_cli_ctx *cli_ctx;
1207 struct vfs_cred vcred = { .vc_uid = 0 };
1213 cli_ctx = ctx_create_kr(sec, &vcred);
1214 if (cli_ctx == NULL)
1217 rc = gss_copy_rvc_cli_ctx(cli_ctx, svc_ctx);
1219 CERROR("failed copy reverse cli ctx: %d\n", rc);
1221 ctx_put_kr(cli_ctx, 1);
1225 rvs_sec_install_root_ctx_kr(sec, cli_ctx, NULL);
1227 ctx_put_kr(cli_ctx, 1);
1232 #else /* ! HAVE_REVERSE_CTX_NOKEY */
1235 int sec_install_rctx_kr(struct ptlrpc_sec *sec,
1236 struct ptlrpc_svc_ctx *svc_ctx)
1238 struct ptlrpc_cli_ctx *cli_ctx = NULL;
1240 struct vfs_cred vcred = { .vc_uid = 0 };
1248 construct_key_desc(desc, sizeof(desc), sec, 0);
1250 key = key_alloc(&gss_key_type, desc, 0, 0,
1251 KEY_POS_ALL | KEY_USR_ALL, 1);
1253 CERROR("failed to alloc key: %ld\n", PTR_ERR(key));
1254 return PTR_ERR(key);
1257 rc = key_instantiate_and_link(key, NULL, 0, NULL, NULL);
1259 CERROR("failed to instantiate key: %d\n", rc);
1263 down_write(&key->sem);
1265 LASSERT(!key_get_payload(key, 0));
1267 cli_ctx = ctx_create_kr(sec, &vcred);
1268 if (cli_ctx == NULL) {
1273 rc = gss_copy_rvc_cli_ctx(cli_ctx, svc_ctx);
1275 CERROR("failed copy reverse cli ctx: %d\n", rc);
1279 rvs_sec_install_root_ctx_kr(sec, cli_ctx, key);
1281 ctx_put_kr(cli_ctx, 1);
1282 up_write(&key->sem);
1291 ctx_put_kr(cli_ctx, 1);
1293 up_write(&key->sem);
1299 #endif /* HAVE_REVERSE_CTX_NOKEY */
1301 /****************************************
1303 ****************************************/
1306 int gss_svc_accept_kr(struct ptlrpc_request *req)
1308 return gss_svc_accept(&gss_policy_keyring, req);
1312 int gss_svc_install_rctx_kr(struct obd_import *imp,
1313 struct ptlrpc_svc_ctx *svc_ctx)
1315 struct ptlrpc_sec *sec;
1318 sec = sptlrpc_import_sec_ref(imp);
1321 rc = sec_install_rctx_kr(sec, svc_ctx);
1322 sptlrpc_sec_put(sec);
1327 /****************************************
1329 ****************************************/
1332 #ifdef HAVE_KEY_TYPE_INSTANTIATE_2ARGS
1333 int gss_kt_instantiate(struct key *key, struct key_preparsed_payload *prep)
1335 const void *data = prep->data;
1336 size_t datalen = prep->datalen;
1338 int gss_kt_instantiate(struct key *key, const void *data, size_t datalen)
1344 if (data != NULL || datalen != 0) {
1345 CERROR("invalid: data %p, len %lu\n", data, (long)datalen);
1349 if (key_get_payload(key, 0)) {
1350 CERROR("key already have payload\n");
1354 /* link the key to session keyring, so following context negotiation
1355 * rpc fired from user space could find this key. This will be unlinked
1356 * automatically when upcall processes die.
1358 * we can't do this through keyctl from userspace, because the upcall
1359 * might be neither possessor nor owner of the key (setuid).
1361 * the session keyring is created upon upcall, and don't change all
1362 * the way until upcall finished, so rcu lock is not needed here.
1364 LASSERT(current_cred()->session_keyring);
1367 rc = key_link(current_cred()->session_keyring, key);
1370 CERROR("failed to link key %08x to keyring %08x: %d\n",
1372 current_cred()->session_keyring->serial, rc);
1376 CDEBUG(D_SEC, "key %p instantiated, ctx %p\n", key,
1377 key_get_payload(key, 0));
1382 * called with key semaphore write locked. it means we can operate
1383 * on the context without fear of loosing refcount.
1386 #ifdef HAVE_KEY_TYPE_INSTANTIATE_2ARGS
1387 int gss_kt_update(struct key *key, struct key_preparsed_payload *prep)
1389 const void *data = prep->data;
1390 __u32 datalen32 = (__u32) prep->datalen;
1392 int gss_kt_update(struct key *key, const void *data, size_t datalen)
1394 __u32 datalen32 = (__u32) datalen;
1396 struct ptlrpc_cli_ctx *ctx = key_get_payload(key, 0);
1397 struct gss_cli_ctx *gctx;
1398 rawobj_t tmpobj = RAWOBJ_EMPTY;
1402 if (data == NULL || datalen32 == 0) {
1403 CWARN("invalid: data %p, len %lu\n", data, (long)datalen32);
1407 /* if upcall finished negotiation too fast (mostly likely because
1408 * of local error happened) and call kt_update(), the ctx
1409 * might be still NULL. but the key will finally be associate
1410 * with a context, or be revoked. if key status is fine, return
1411 * -EAGAIN to allow userspace sleep a while and call again. */
1413 CDEBUG(D_SEC, "update too soon: key %p(%x) flags %lx\n",
1414 key, key->serial, key->flags);
1416 rc = key_validate(key);
1423 LASSERT(atomic_read(&ctx->cc_refcount) > 0);
1424 LASSERT(ctx->cc_sec);
1426 ctx_clear_timer_kr(ctx);
1428 /* don't proceed if already refreshed */
1429 if (cli_ctx_is_refreshed(ctx)) {
1430 CWARN("ctx already done refresh\n");
1434 sptlrpc_cli_ctx_get(ctx);
1435 gctx = ctx2gctx(ctx);
1437 rc = buffer_extract_bytes(&data, &datalen32, &gctx->gc_win,
1438 sizeof(gctx->gc_win));
1440 CERROR("failed extract seq_win\n");
1444 if (gctx->gc_win == 0) {
1445 __u32 nego_rpc_err, nego_gss_err;
1447 rc = buffer_extract_bytes(&data, &datalen32, &nego_rpc_err,
1448 sizeof(nego_rpc_err));
1450 CERROR("cannot extract RPC: rc = %d\n", rc);
1454 rc = buffer_extract_bytes(&data, &datalen32, &nego_gss_err,
1455 sizeof(nego_gss_err));
1457 CERROR("failed to extract gss rc = %d\n", rc);
1461 CERROR("negotiation: rpc err %d, gss err %x\n",
1462 nego_rpc_err, nego_gss_err);
1464 rc = nego_rpc_err ? nego_rpc_err : -EACCES;
1466 rc = rawobj_extract_local_alloc(&gctx->gc_handle,
1467 (__u32 **) &data, &datalen32);
1469 CERROR("failed extract handle\n");
1473 rc = rawobj_extract_local(&tmpobj,
1474 (__u32 **) &data, &datalen32);
1476 CERROR("failed extract mech\n");
1480 rc = lgss_import_sec_context(&tmpobj,
1481 sec2gsec(ctx->cc_sec)->gs_mech,
1483 if (rc != GSS_S_COMPLETE)
1484 CERROR("failed import context\n");
1489 /* we don't care what current status of this ctx, even someone else
1490 * is operating on the ctx at the same time. we just add up our own
1493 gss_cli_ctx_uptodate(gctx);
1495 /* this will also revoke the key. has to be done before
1496 * wakeup waiters otherwise they can find the stale key */
1497 kill_key_locked(key);
1499 cli_ctx_expire(ctx);
1501 if (rc != -ERESTART)
1502 set_bit(PTLRPC_CTX_ERROR_BIT, &ctx->cc_flags);
1505 /* let user space think it's a success */
1506 sptlrpc_cli_ctx_put(ctx, 1);
1510 #ifndef HAVE_KEY_MATCH_DATA
1512 gss_kt_match(const struct key *key, const void *desc)
1514 return strcmp(key->description, (const char *) desc) == 0 &&
1515 !test_bit(KEY_FLAG_REVOKED, &key->flags);
1517 #else /* ! HAVE_KEY_MATCH_DATA */
1519 gss_kt_match(const struct key *key, const struct key_match_data *match_data)
1521 const char *desc = match_data->raw_data;
1523 return strcmp(key->description, desc) == 0 &&
1524 !test_bit(KEY_FLAG_REVOKED, &key->flags);
1528 * Preparse the match criterion.
1530 static int gss_kt_match_preparse(struct key_match_data *match_data)
1532 match_data->lookup_type = KEYRING_SEARCH_LOOKUP_DIRECT;
1533 match_data->cmp = gss_kt_match;
1536 #endif /* HAVE_KEY_MATCH_DATA */
1539 void gss_kt_destroy(struct key *key)
1542 LASSERT(!key_get_payload(key, 0));
1543 CDEBUG(D_SEC, "destroy key %p\n", key);
1548 void gss_kt_describe(const struct key *key, struct seq_file *s)
1550 if (key->description == NULL)
1551 seq_puts(s, "[null]");
1553 seq_puts(s, key->description);
1556 static struct key_type gss_key_type =
1560 .instantiate = gss_kt_instantiate,
1561 .update = gss_kt_update,
1562 #ifdef HAVE_KEY_MATCH_DATA
1563 .match_preparse = gss_kt_match_preparse,
1565 .match = gss_kt_match,
1567 .destroy = gss_kt_destroy,
1568 .describe = gss_kt_describe,
1571 /****************************************
1572 * lustre gss keyring policy *
1573 ****************************************/
1575 static struct ptlrpc_ctx_ops gss_keyring_ctxops = {
1576 .match = gss_cli_ctx_match,
1577 .refresh = gss_cli_ctx_refresh_kr,
1578 .validate = gss_cli_ctx_validate_kr,
1579 .die = gss_cli_ctx_die_kr,
1580 .sign = gss_cli_ctx_sign,
1581 .verify = gss_cli_ctx_verify,
1582 .seal = gss_cli_ctx_seal,
1583 .unseal = gss_cli_ctx_unseal,
1584 .wrap_bulk = gss_cli_ctx_wrap_bulk,
1585 .unwrap_bulk = gss_cli_ctx_unwrap_bulk,
1588 static struct ptlrpc_sec_cops gss_sec_keyring_cops = {
1589 .create_sec = gss_sec_create_kr,
1590 .destroy_sec = gss_sec_destroy_kr,
1591 .kill_sec = gss_sec_kill,
1592 .lookup_ctx = gss_sec_lookup_ctx_kr,
1593 .release_ctx = gss_sec_release_ctx_kr,
1594 .flush_ctx_cache = gss_sec_flush_ctx_cache_kr,
1595 .gc_ctx = gss_sec_gc_ctx_kr,
1596 .install_rctx = gss_sec_install_rctx,
1597 .alloc_reqbuf = gss_alloc_reqbuf,
1598 .free_reqbuf = gss_free_reqbuf,
1599 .alloc_repbuf = gss_alloc_repbuf,
1600 .free_repbuf = gss_free_repbuf,
1601 .enlarge_reqbuf = gss_enlarge_reqbuf,
1602 .display = gss_sec_display_kr,
1605 static struct ptlrpc_sec_sops gss_sec_keyring_sops = {
1606 .accept = gss_svc_accept_kr,
1607 .invalidate_ctx = gss_svc_invalidate_ctx,
1608 .alloc_rs = gss_svc_alloc_rs,
1609 .authorize = gss_svc_authorize,
1610 .free_rs = gss_svc_free_rs,
1611 .free_ctx = gss_svc_free_ctx,
1612 .prep_bulk = gss_svc_prep_bulk,
1613 .unwrap_bulk = gss_svc_unwrap_bulk,
1614 .wrap_bulk = gss_svc_wrap_bulk,
1615 .install_rctx = gss_svc_install_rctx_kr,
1618 static struct ptlrpc_sec_policy gss_policy_keyring = {
1619 .sp_owner = THIS_MODULE,
1620 .sp_name = "gss.keyring",
1621 .sp_policy = SPTLRPC_POLICY_GSS,
1622 .sp_cops = &gss_sec_keyring_cops,
1623 .sp_sops = &gss_sec_keyring_sops,
1627 int __init gss_init_keyring(void)
1631 rc = register_key_type(&gss_key_type);
1633 CERROR("failed to register keyring type: %d\n", rc);
1637 rc = sptlrpc_register_policy(&gss_policy_keyring);
1639 unregister_key_type(&gss_key_type);
1646 void __exit gss_exit_keyring(void)
1648 unregister_key_type(&gss_key_type);
1649 sptlrpc_unregister_policy(&gss_policy_keyring);