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/
30 * Lustre is a trademark of Sun Microsystems, Inc.
32 * lustre/ptlrpc/gss/gss_keyring.c
34 * Author: Eric Mei <ericm@clusterfs.com>
37 #define DEBUG_SUBSYSTEM S_SEC
38 #include <linux/init.h>
39 #include <linux/module.h>
40 #include <linux/slab.h>
41 #include <linux/dcache.h>
43 #include <linux/crypto.h>
44 #include <linux/key.h>
45 #include <linux/keyctl.h>
46 #include <linux/key-type.h>
47 #include <linux/mutex.h>
48 #include <asm/atomic.h>
50 #include <libcfs/linux/linux-list.h>
52 #include <obd_class.h>
53 #include <obd_support.h>
54 #include <uapi/linux/lustre/lustre_idl.h>
55 #include <lustre_sec.h>
56 #include <lustre_net.h>
57 #include <lustre_import.h>
60 #include "gss_internal.h"
63 static struct ptlrpc_sec_policy gss_policy_keyring;
64 static struct ptlrpc_ctx_ops gss_keyring_ctxops;
65 static struct key_type gss_key_type;
67 static int sec_install_rctx_kr(struct ptlrpc_sec *sec,
68 struct ptlrpc_svc_ctx *svc_ctx);
71 * the timeout is only for the case that upcall child process die abnormally.
72 * in any other cases it should finally update kernel key.
74 * FIXME we'd better to incorporate the client & server side upcall timeouts
75 * into the framework of Adaptive Timeouts, but we need to figure out how to
76 * make sure that kernel knows the upcall processes is in-progress or died
79 #define KEYRING_UPCALL_TIMEOUT (obd_timeout + obd_timeout)
81 /****************************************
83 ****************************************/
85 #define DUMP_PROCESS_KEYRINGS(tsk) \
87 CWARN("DUMP PK: %s[%u,%u/%u](<-%s[%u,%u/%u]): " \
88 "a %d, t %d, p %d, s %d, u %d, us %d, df %d\n", \
89 tsk->comm, tsk->pid, tsk->uid, tsk->fsuid, \
90 tsk->parent->comm, tsk->parent->pid, \
91 tsk->parent->uid, tsk->parent->fsuid, \
92 tsk->request_key_auth ? \
93 tsk->request_key_auth->serial : 0, \
94 key_cred(tsk)->thread_keyring ? \
95 key_cred(tsk)->thread_keyring->serial : 0, \
96 key_tgcred(tsk)->process_keyring ? \
97 key_tgcred(tsk)->process_keyring->serial : 0, \
98 key_tgcred(tsk)->session_keyring ? \
99 key_tgcred(tsk)->session_keyring->serial : 0, \
100 key_cred(tsk)->user->uid_keyring ? \
101 key_cred(tsk)->user->uid_keyring->serial : 0, \
102 key_cred(tsk)->user->session_keyring ? \
103 key_cred(tsk)->user->session_keyring->serial : 0, \
104 key_cred(tsk)->jit_keyring \
108 #define DUMP_KEY(key) \
110 CWARN("DUMP KEY: %p(%d) ref %d u%u/g%u desc %s\n", \
111 key, key->serial, ll_read_key_usage(key), \
112 key->uid, key->gid, \
113 key->description ? key->description : "n/a" \
117 #define key_cred(tsk) ((tsk)->cred)
118 #ifdef HAVE_CRED_TGCRED
119 #define key_tgcred(tsk) ((tsk)->cred->tgcred)
121 #define key_tgcred(tsk) key_cred(tsk)
124 static inline void keyring_upcall_lock(struct gss_sec_keyring *gsec_kr)
126 #ifdef HAVE_KEYRING_UPCALL_SERIALIZED
127 mutex_lock(&gsec_kr->gsk_uc_lock);
131 static inline void keyring_upcall_unlock(struct gss_sec_keyring *gsec_kr)
133 #ifdef HAVE_KEYRING_UPCALL_SERIALIZED
134 mutex_unlock(&gsec_kr->gsk_uc_lock);
138 static inline void key_revoke_locked(struct key *key)
140 set_bit(KEY_FLAG_REVOKED, &key->flags);
143 static void ctx_upcall_timeout_kr(cfs_timer_cb_arg_t data)
145 struct gss_cli_ctx_keyring *gctx_kr = cfs_from_timer(gctx_kr,
147 struct ptlrpc_cli_ctx *ctx = &(gctx_kr->gck_base.gc_base);
148 struct key *key = gctx_kr->gck_key;
150 CWARN("ctx %p, key %p\n", ctx, key);
155 key_revoke_locked(key);
158 static void ctx_start_timer_kr(struct ptlrpc_cli_ctx *ctx, time64_t timeout)
160 struct gss_cli_ctx_keyring *gctx_kr = ctx2gctx_keyring(ctx);
161 struct timer_list *timer = gctx_kr->gck_timer;
165 CDEBUG(D_SEC, "ctx %p: start timer %llds\n", ctx, timeout);
167 cfs_timer_setup(timer, ctx_upcall_timeout_kr,
168 (unsigned long)gctx_kr, 0);
169 timer->expires = cfs_time_seconds(timeout) + jiffies;
174 * caller should make sure no race with other threads
177 void ctx_clear_timer_kr(struct ptlrpc_cli_ctx *ctx)
179 struct gss_cli_ctx_keyring *gctx_kr = ctx2gctx_keyring(ctx);
180 struct timer_list *timer = gctx_kr->gck_timer;
185 CDEBUG(D_SEC, "ctx %p, key %p\n", ctx, gctx_kr->gck_key);
187 gctx_kr->gck_timer = NULL;
189 del_singleshot_timer_sync(timer);
195 struct ptlrpc_cli_ctx *ctx_create_kr(struct ptlrpc_sec *sec,
196 struct vfs_cred *vcred)
198 struct ptlrpc_cli_ctx *ctx;
199 struct gss_cli_ctx_keyring *gctx_kr;
201 OBD_ALLOC_PTR(gctx_kr);
205 OBD_ALLOC_PTR(gctx_kr->gck_timer);
206 if (gctx_kr->gck_timer == NULL) {
207 OBD_FREE_PTR(gctx_kr);
210 cfs_timer_setup(gctx_kr->gck_timer, NULL, 0, 0);
212 ctx = &gctx_kr->gck_base.gc_base;
214 if (gss_cli_ctx_init_common(sec, ctx, &gss_keyring_ctxops, vcred)) {
215 OBD_FREE_PTR(gctx_kr->gck_timer);
216 OBD_FREE_PTR(gctx_kr);
220 ctx->cc_expire = ktime_get_real_seconds() + KEYRING_UPCALL_TIMEOUT;
221 clear_bit(PTLRPC_CTX_NEW_BIT, &ctx->cc_flags);
222 atomic_inc(&ctx->cc_refcount); /* for the caller */
227 static void ctx_destroy_kr(struct ptlrpc_cli_ctx *ctx)
229 struct ptlrpc_sec *sec = ctx->cc_sec;
230 struct gss_cli_ctx_keyring *gctx_kr = ctx2gctx_keyring(ctx);
232 CDEBUG(D_SEC, "destroying ctx %p\n", ctx);
234 /* at this time the association with key has been broken. */
236 LASSERT(atomic_read(&sec->ps_refcount) > 0);
237 LASSERT(atomic_read(&sec->ps_nctx) > 0);
238 LASSERT(test_bit(PTLRPC_CTX_CACHED_BIT, &ctx->cc_flags) == 0);
239 LASSERT(gctx_kr->gck_key == NULL);
241 ctx_clear_timer_kr(ctx);
242 LASSERT(gctx_kr->gck_timer == NULL);
244 if (gss_cli_ctx_fini_common(sec, ctx))
247 OBD_FREE_PTR(gctx_kr);
249 atomic_dec(&sec->ps_nctx);
250 sptlrpc_sec_put(sec);
253 static void ctx_release_kr(struct ptlrpc_cli_ctx *ctx, int sync)
258 atomic_inc(&ctx->cc_refcount);
259 sptlrpc_gc_add_ctx(ctx);
263 static void ctx_put_kr(struct ptlrpc_cli_ctx *ctx, int sync)
265 LASSERT(atomic_read(&ctx->cc_refcount) > 0);
267 if (atomic_dec_and_test(&ctx->cc_refcount))
268 ctx_release_kr(ctx, sync);
272 * key <-> ctx association and rules:
273 * - ctx might not bind with any key
274 * - key/ctx binding is protected by key semaphore (if the key present)
275 * - key and ctx each take a reference of the other
276 * - ctx enlist/unlist is protected by ctx spinlock
277 * - never enlist a ctx after it's been unlisted
278 * - whoever do enlist should also do bind, lock key before enlist:
279 * - lock key -> lock ctx -> enlist -> unlock ctx -> bind -> unlock key
280 * - whoever do unlist should also do unbind:
281 * - lock key -> lock ctx -> unlist -> unlock ctx -> unbind -> unlock key
282 * - lock ctx -> unlist -> unlock ctx -> lock key -> unbind -> unlock key
285 static inline void spin_lock_if(spinlock_t *lock, int condition)
291 static inline void spin_unlock_if(spinlock_t *lock, int condition)
297 static void ctx_enlist_kr(struct ptlrpc_cli_ctx *ctx, int is_root, int locked)
299 struct ptlrpc_sec *sec = ctx->cc_sec;
300 struct gss_sec_keyring *gsec_kr = sec2gsec_keyring(sec);
302 LASSERT(!test_bit(PTLRPC_CTX_CACHED_BIT, &ctx->cc_flags));
303 LASSERT(atomic_read(&ctx->cc_refcount) > 0);
305 spin_lock_if(&sec->ps_lock, !locked);
307 atomic_inc(&ctx->cc_refcount);
308 set_bit(PTLRPC_CTX_CACHED_BIT, &ctx->cc_flags);
309 hlist_add_head(&ctx->cc_cache, &gsec_kr->gsk_clist);
311 gsec_kr->gsk_root_ctx = ctx;
313 spin_unlock_if(&sec->ps_lock, !locked);
317 * Note after this get called, caller should not access ctx again because
318 * it might have been freed, unless caller hold at least one refcount of
321 * return non-zero if we indeed unlist this ctx.
323 static int ctx_unlist_kr(struct ptlrpc_cli_ctx *ctx, int locked)
325 struct ptlrpc_sec *sec = ctx->cc_sec;
326 struct gss_sec_keyring *gsec_kr = sec2gsec_keyring(sec);
328 /* if hashed bit has gone, leave the job to somebody who is doing it */
329 if (test_and_clear_bit(PTLRPC_CTX_CACHED_BIT, &ctx->cc_flags) == 0)
332 /* drop ref inside spin lock to prevent race with other operations */
333 spin_lock_if(&sec->ps_lock, !locked);
335 if (gsec_kr->gsk_root_ctx == ctx)
336 gsec_kr->gsk_root_ctx = NULL;
337 hlist_del_init(&ctx->cc_cache);
338 atomic_dec(&ctx->cc_refcount);
340 spin_unlock_if(&sec->ps_lock, !locked);
346 * Get specific payload. Newer kernels support 4 slots.
349 key_get_payload(struct key *key, unsigned int index)
351 void *key_ptr = NULL;
353 #ifdef HAVE_KEY_PAYLOAD_DATA_ARRAY
354 key_ptr = key->payload.data[index];
357 key_ptr = key->payload.data;
363 * Set specific payload. Newer kernels support 4 slots.
365 static int key_set_payload(struct key *key, unsigned int index,
366 struct ptlrpc_cli_ctx *ctx)
370 #ifdef HAVE_KEY_PAYLOAD_DATA_ARRAY
372 key->payload.data[index] = ctx;
375 key->payload.data = ctx;
383 * bind a key with a ctx together.
384 * caller must hold write lock of the key, as well as ref on key & ctx.
386 static void bind_key_ctx(struct key *key, struct ptlrpc_cli_ctx *ctx)
388 LASSERT(atomic_read(&ctx->cc_refcount) > 0);
389 LASSERT(ll_read_key_usage(key) > 0);
390 LASSERT(ctx2gctx_keyring(ctx)->gck_key == NULL);
391 LASSERT(!key_get_payload(key, 0));
393 /* at this time context may or may not in list. */
395 atomic_inc(&ctx->cc_refcount);
396 ctx2gctx_keyring(ctx)->gck_key = key;
397 LASSERT(!key_set_payload(key, 0, ctx));
401 * unbind a key and a ctx.
402 * caller must hold write lock, as well as a ref of the key.
404 static void unbind_key_ctx(struct key *key, struct ptlrpc_cli_ctx *ctx)
406 LASSERT(key_get_payload(key, 0) == ctx);
407 LASSERT(test_bit(PTLRPC_CTX_CACHED_BIT, &ctx->cc_flags) == 0);
409 /* must revoke the key, or others may treat it as newly created */
410 key_revoke_locked(key);
412 key_set_payload(key, 0, NULL);
413 ctx2gctx_keyring(ctx)->gck_key = NULL;
415 /* once ctx get split from key, the timer is meaningless */
416 ctx_clear_timer_kr(ctx);
423 * given a ctx, unbind with its coupled key, if any.
424 * unbind could only be called once, so we don't worry the key be released
427 static void unbind_ctx_kr(struct ptlrpc_cli_ctx *ctx)
429 struct key *key = ctx2gctx_keyring(ctx)->gck_key;
432 LASSERT(key_get_payload(key, 0) == ctx);
435 down_write(&key->sem);
436 unbind_key_ctx(key, ctx);
443 * given a key, unbind with its coupled ctx, if any.
444 * caller must hold write lock, as well as a ref of the key.
446 static void unbind_key_locked(struct key *key)
448 struct ptlrpc_cli_ctx *ctx = key_get_payload(key, 0);
451 unbind_key_ctx(key, ctx);
455 * unlist a ctx, and unbind from coupled key
457 static void kill_ctx_kr(struct ptlrpc_cli_ctx *ctx)
459 if (ctx_unlist_kr(ctx, 0))
464 * given a key, unlist and unbind with the coupled ctx (if any).
465 * caller must hold write lock, as well as a ref of the key.
467 static void kill_key_locked(struct key *key)
469 struct ptlrpc_cli_ctx *ctx = key_get_payload(key, 0);
471 if (ctx && ctx_unlist_kr(ctx, 0))
472 unbind_key_locked(key);
476 * caller should hold one ref on contexts in freelist.
478 static void dispose_ctx_list_kr(struct hlist_head *freelist)
480 struct hlist_node __maybe_unused *pos, *next;
481 struct ptlrpc_cli_ctx *ctx;
482 struct gss_cli_ctx *gctx;
484 cfs_hlist_for_each_entry_safe(ctx, pos, next, freelist, cc_cache) {
485 hlist_del_init(&ctx->cc_cache);
487 /* reverse ctx: update current seq to buddy svcctx if exist.
488 * ideally this should be done at gss_cli_ctx_finalize(), but
489 * the ctx destroy could be delayed by:
490 * 1) ctx still has reference;
491 * 2) ctx destroy is asynchronous;
492 * and reverse import call inval_all_ctx() require this be done
493 * _immediately_ otherwise newly created reverse ctx might copy
494 * the very old sequence number from svcctx. */
495 gctx = ctx2gctx(ctx);
496 if (!rawobj_empty(&gctx->gc_svc_handle) &&
497 sec_is_reverse(gctx->gc_base.cc_sec)) {
498 gss_svc_upcall_update_sequence(&gctx->gc_svc_handle,
499 (__u32) atomic_read(&gctx->gc_seq));
502 /* we need to wakeup waiting reqs here. the context might
503 * be forced released before upcall finished, then the
504 * late-arrived downcall can't find the ctx even. */
505 sptlrpc_cli_ctx_wakeup(ctx);
513 * lookup a root context directly in a sec, return root ctx with a
514 * reference taken or NULL.
517 struct ptlrpc_cli_ctx * sec_lookup_root_ctx_kr(struct ptlrpc_sec *sec)
519 struct gss_sec_keyring *gsec_kr = sec2gsec_keyring(sec);
520 struct ptlrpc_cli_ctx *ctx = NULL;
522 spin_lock(&sec->ps_lock);
524 ctx = gsec_kr->gsk_root_ctx;
526 if (ctx == NULL && unlikely(sec_is_reverse(sec))) {
527 struct hlist_node __maybe_unused *node;
528 struct ptlrpc_cli_ctx *tmp;
530 /* reverse ctx, search root ctx in list, choose the one
531 * with shortest expire time, which is most possibly have
532 * an established peer ctx at client side. */
533 cfs_hlist_for_each_entry(tmp, node, &gsec_kr->gsk_clist,
535 if (ctx == NULL || ctx->cc_expire == 0 ||
536 ctx->cc_expire > tmp->cc_expire) {
538 /* promote to be root_ctx */
539 gsec_kr->gsk_root_ctx = ctx;
545 LASSERT(atomic_read(&ctx->cc_refcount) > 0);
546 LASSERT(!hlist_empty(&gsec_kr->gsk_clist));
547 atomic_inc(&ctx->cc_refcount);
550 spin_unlock(&sec->ps_lock);
555 #define RVS_CTX_EXPIRE_NICE (10)
558 void rvs_sec_install_root_ctx_kr(struct ptlrpc_sec *sec,
559 struct ptlrpc_cli_ctx *new_ctx,
562 struct gss_sec_keyring *gsec_kr = sec2gsec_keyring(sec);
563 struct hlist_node __maybe_unused *hnode;
564 struct ptlrpc_cli_ctx *ctx;
568 LASSERT(sec_is_reverse(sec));
570 spin_lock(&sec->ps_lock);
572 now = ktime_get_real_seconds();
574 /* set all existing ctxs short expiry */
575 cfs_hlist_for_each_entry(ctx, hnode, &gsec_kr->gsk_clist, cc_cache) {
576 if (ctx->cc_expire > now + RVS_CTX_EXPIRE_NICE) {
577 ctx->cc_early_expire = 1;
578 ctx->cc_expire = now + RVS_CTX_EXPIRE_NICE;
582 /* if there's root_ctx there, instead obsolete the current
583 * immediately, we leave it continue operating for a little while.
584 * hopefully when the first backward rpc with newest ctx send out,
585 * the client side already have the peer ctx well established. */
586 ctx_enlist_kr(new_ctx, gsec_kr->gsk_root_ctx ? 0 : 1, 1);
589 bind_key_ctx(key, new_ctx);
591 spin_unlock(&sec->ps_lock);
594 static void construct_key_desc(void *buf, int bufsize,
595 struct ptlrpc_sec *sec, uid_t uid)
597 snprintf(buf, bufsize, "%d@%x", uid, sec->ps_id);
598 ((char *)buf)[bufsize - 1] = '\0';
601 /****************************************
603 ****************************************/
606 struct ptlrpc_sec * gss_sec_create_kr(struct obd_import *imp,
607 struct ptlrpc_svc_ctx *svcctx,
608 struct sptlrpc_flavor *sf)
610 struct gss_sec_keyring *gsec_kr;
613 OBD_ALLOC(gsec_kr, sizeof(*gsec_kr));
617 INIT_HLIST_HEAD(&gsec_kr->gsk_clist);
618 gsec_kr->gsk_root_ctx = NULL;
619 mutex_init(&gsec_kr->gsk_root_uc_lock);
620 #ifdef HAVE_KEYRING_UPCALL_SERIALIZED
621 mutex_init(&gsec_kr->gsk_uc_lock);
624 if (gss_sec_create_common(&gsec_kr->gsk_base, &gss_policy_keyring,
628 if (svcctx != NULL &&
629 sec_install_rctx_kr(&gsec_kr->gsk_base.gs_base, svcctx)) {
630 gss_sec_destroy_common(&gsec_kr->gsk_base);
634 RETURN(&gsec_kr->gsk_base.gs_base);
637 OBD_FREE(gsec_kr, sizeof(*gsec_kr));
642 void gss_sec_destroy_kr(struct ptlrpc_sec *sec)
644 struct gss_sec *gsec = sec2gsec(sec);
645 struct gss_sec_keyring *gsec_kr = sec2gsec_keyring(sec);
647 CDEBUG(D_SEC, "destroy %s@%p\n", sec->ps_policy->sp_name, sec);
649 LASSERT(hlist_empty(&gsec_kr->gsk_clist));
650 LASSERT(gsec_kr->gsk_root_ctx == NULL);
652 gss_sec_destroy_common(gsec);
654 OBD_FREE(gsec_kr, sizeof(*gsec_kr));
657 static inline int user_is_root(struct ptlrpc_sec *sec, struct vfs_cred *vcred)
659 /* except the ROOTONLY flag, treat it as root user only if real uid
660 * is 0, euid/fsuid being 0 are handled as setuid scenarios */
661 if (sec_is_rootonly(sec) || (vcred->vc_uid == 0))
668 * unlink request key from it's ring, which is linked during request_key().
669 * sadly, we have to 'guess' which keyring it's linked to.
671 * FIXME this code is fragile, depend on how request_key_link() is implemented.
673 static void request_key_unlink(struct key *key)
675 struct task_struct *tsk = current;
678 switch (key_cred(tsk)->jit_keyring) {
679 case KEY_REQKEY_DEFL_DEFAULT:
680 case KEY_REQKEY_DEFL_THREAD_KEYRING:
681 ring = key_get(key_cred(tsk)->thread_keyring);
685 case KEY_REQKEY_DEFL_PROCESS_KEYRING:
686 ring = key_get(key_tgcred(tsk)->process_keyring);
690 case KEY_REQKEY_DEFL_SESSION_KEYRING:
692 ring = key_get(rcu_dereference(key_tgcred(tsk)
698 case KEY_REQKEY_DEFL_USER_SESSION_KEYRING:
699 ring = key_get(key_cred(tsk)->user->session_keyring);
701 case KEY_REQKEY_DEFL_USER_KEYRING:
702 ring = key_get(key_cred(tsk)->user->uid_keyring);
704 case KEY_REQKEY_DEFL_GROUP_KEYRING:
710 key_unlink(ring, key);
715 struct ptlrpc_cli_ctx * gss_sec_lookup_ctx_kr(struct ptlrpc_sec *sec,
716 struct vfs_cred *vcred,
717 int create, int remove_dead)
719 struct obd_import *imp = sec->ps_import;
720 struct gss_sec_keyring *gsec_kr = sec2gsec_keyring(sec);
721 struct ptlrpc_cli_ctx *ctx = NULL;
722 unsigned int is_root = 0, create_new = 0;
727 const char *sec_part_flags = "";
731 LASSERT(imp != NULL);
733 is_root = user_is_root(sec, vcred);
735 /* a little bit optimization for root context */
737 ctx = sec_lookup_root_ctx_kr(sec);
739 * Only lookup directly for REVERSE sec, which should
742 if (ctx || sec_is_reverse(sec))
746 LASSERT(create != 0);
748 /* for root context, obtain lock and check again, this time hold
749 * the root upcall lock, make sure nobody else populated new root
750 * context after last check. */
752 mutex_lock(&gsec_kr->gsk_root_uc_lock);
754 ctx = sec_lookup_root_ctx_kr(sec);
758 /* update reverse handle for root user */
759 sec2gsec(sec)->gs_rvs_hdl = gss_get_next_ctx_index();
761 switch (sec->ps_part) {
763 sec_part_flags = "m";
766 sec_part_flags = "o";
769 sec_part_flags = "rmo";
772 sec_part_flags = "r";
779 switch (SPTLRPC_FLVR_SVC(sec->ps_flvr.sf_rpc)) {
780 case SPTLRPC_SVC_NULL:
783 case SPTLRPC_SVC_AUTH:
786 case SPTLRPC_SVC_INTG:
789 case SPTLRPC_SVC_PRIV:
797 /* in case of setuid, key will be constructed as owner of fsuid/fsgid,
798 * but we do authentication based on real uid/gid. the key permission
799 * bits will be exactly as POS_ALL, so only processes who subscribed
800 * this key could have the access, although the quota might be counted
801 * on others (fsuid/fsgid).
803 * keyring will use fsuid/fsgid as upcall parameters, so we have to
804 * encode real uid/gid into callout info.
807 /* But first we need to make sure the obd type is supported */
808 if (strcmp(imp->imp_obd->obd_type->typ_name, LUSTRE_MDC_NAME) &&
809 strcmp(imp->imp_obd->obd_type->typ_name, LUSTRE_OSC_NAME) &&
810 strcmp(imp->imp_obd->obd_type->typ_name, LUSTRE_MGC_NAME) &&
811 strcmp(imp->imp_obd->obd_type->typ_name, LUSTRE_LWP_NAME) &&
812 strcmp(imp->imp_obd->obd_type->typ_name, LUSTRE_OSP_NAME)) {
813 CERROR("obd %s is not a supported device\n",
814 imp->imp_obd->obd_name);
815 GOTO(out, ctx = NULL);
818 construct_key_desc(desc, sizeof(desc), sec, vcred->vc_uid);
820 /* callout info format:
821 * secid:mech:uid:gid:sec_flags:svc_flag:svc_type:peer_nid:target_uuid:
824 coinfo_size = sizeof(struct obd_uuid) + MAX_OBD_NAME + 64;
825 OBD_ALLOC(coinfo, coinfo_size);
829 /* Last callout parameter is pid of process whose namespace will be used
830 * for credentials' retrieval.
831 * For user's credentials (in which case sec_part_flags is empty), use
832 * current PID instead of import's reference PID to get reference
834 snprintf(coinfo, coinfo_size, "%d:%s:%u:%u:%s:%c:%d:%#llx:%s:%#llx:%d",
835 sec->ps_id, sec2gsec(sec)->gs_mech->gm_name,
836 vcred->vc_uid, vcred->vc_gid,
837 sec_part_flags, svc_flag, import_to_gss_svc(imp),
838 imp->imp_connection->c_peer.nid, imp->imp_obd->obd_name,
839 imp->imp_connection->c_self,
840 sec_part_flags[0] == '\0' ?
841 current_pid() : imp->imp_sec_refpid);
843 CDEBUG(D_SEC, "requesting key for %s\n", desc);
845 keyring_upcall_lock(gsec_kr);
846 key = request_key(&gss_key_type, desc, coinfo);
847 keyring_upcall_unlock(gsec_kr);
849 OBD_FREE(coinfo, coinfo_size);
852 CERROR("failed request key: %ld\n", PTR_ERR(key));
855 CDEBUG(D_SEC, "obtained key %08x for %s\n", key->serial, desc);
857 /* once payload.data was pointed to a ctx, it never changes until
858 * we de-associate them; but parallel request_key() may return
859 * a key with payload.data == NULL at the same time. so we still
860 * need wirtelock of key->sem to serialize them. */
861 down_write(&key->sem);
863 ctx = key_get_payload(key, 0);
865 LASSERT(atomic_read(&ctx->cc_refcount) >= 1);
866 LASSERT(ctx2gctx_keyring(ctx)->gck_key == key);
867 LASSERT(ll_read_key_usage(key) >= 2);
869 /* simply take a ref and return. it's upper layer's
870 * responsibility to detect & replace dead ctx. */
871 atomic_inc(&ctx->cc_refcount);
873 /* pre initialization with a cli_ctx. this can't be done in
874 * key_instantiate() because we'v no enough information
876 ctx = ctx_create_kr(sec, vcred);
878 ctx_enlist_kr(ctx, is_root, 0);
879 bind_key_ctx(key, ctx);
881 ctx_start_timer_kr(ctx, KEYRING_UPCALL_TIMEOUT);
883 CDEBUG(D_SEC, "installed key %p <-> ctx %p (sec %p)\n",
886 /* we'd prefer to call key_revoke(), but we more like
887 * to revoke it within this key->sem locked period. */
888 key_revoke_locked(key);
896 if (is_root && create_new)
897 request_key_unlink(key);
902 mutex_unlock(&gsec_kr->gsk_root_uc_lock);
907 void gss_sec_release_ctx_kr(struct ptlrpc_sec *sec,
908 struct ptlrpc_cli_ctx *ctx,
911 LASSERT(atomic_read(&sec->ps_refcount) > 0);
912 LASSERT(atomic_read(&ctx->cc_refcount) == 0);
913 ctx_release_kr(ctx, sync);
917 * flush context of normal user, we must resort to keyring itself to find out
918 * contexts which belong to me.
920 * Note here we suppose only to flush _my_ context, the "uid" will
921 * be ignored in the search.
924 void flush_user_ctx_cache_kr(struct ptlrpc_sec *sec,
926 int grace, int force)
931 /* nothing to do for reverse or rootonly sec */
932 if (sec_is_reverse(sec) || sec_is_rootonly(sec))
935 construct_key_desc(desc, sizeof(desc), sec, uid);
937 /* there should be only one valid key, but we put it in the
938 * loop in case of any weird cases */
940 key = request_key(&gss_key_type, desc, NULL);
942 CDEBUG(D_SEC, "No more key found for current user\n");
946 down_write(&key->sem);
948 kill_key_locked(key);
950 /* kill_key_locked() should usually revoke the key, but we
951 * revoke it again to make sure, e.g. some case the key may
952 * not well coupled with a context. */
953 key_revoke_locked(key);
957 request_key_unlink(key);
964 * flush context of root or all, we iterate through the list.
967 void flush_spec_ctx_cache_kr(struct ptlrpc_sec *sec, uid_t uid, int grace,
970 struct gss_sec_keyring *gsec_kr;
971 struct hlist_head freelist = HLIST_HEAD_INIT;
972 struct hlist_node __maybe_unused *pos, *next;
973 struct ptlrpc_cli_ctx *ctx;
976 gsec_kr = sec2gsec_keyring(sec);
978 spin_lock(&sec->ps_lock);
979 cfs_hlist_for_each_entry_safe(ctx, pos, next,
980 &gsec_kr->gsk_clist, cc_cache) {
981 LASSERT(atomic_read(&ctx->cc_refcount) > 0);
983 if (uid != -1 && uid != ctx->cc_vcred.vc_uid)
986 /* at this moment there's at least 2 base reference:
987 * key association and in-list. */
988 if (atomic_read(&ctx->cc_refcount) > 2) {
991 CWARN("flush busy ctx %p(%u->%s, extra ref %d)\n",
992 ctx, ctx->cc_vcred.vc_uid,
993 sec2target_str(ctx->cc_sec),
994 atomic_read(&ctx->cc_refcount) - 2);
997 set_bit(PTLRPC_CTX_DEAD_BIT, &ctx->cc_flags);
999 clear_bit(PTLRPC_CTX_UPTODATE_BIT, &ctx->cc_flags);
1001 atomic_inc(&ctx->cc_refcount);
1003 if (ctx_unlist_kr(ctx, 1)) {
1004 hlist_add_head(&ctx->cc_cache, &freelist);
1006 LASSERT(atomic_read(&ctx->cc_refcount) >= 2);
1007 atomic_dec(&ctx->cc_refcount);
1010 spin_unlock(&sec->ps_lock);
1012 dispose_ctx_list_kr(&freelist);
1017 int gss_sec_flush_ctx_cache_kr(struct ptlrpc_sec *sec,
1018 uid_t uid, int grace, int force)
1022 CDEBUG(D_SEC, "sec %p(%d, nctx %d), uid %d, grace %d, force %d\n",
1023 sec, atomic_read(&sec->ps_refcount),
1024 atomic_read(&sec->ps_nctx),
1027 if (uid != -1 && uid != 0)
1028 flush_user_ctx_cache_kr(sec, uid, grace, force);
1030 flush_spec_ctx_cache_kr(sec, uid, grace, force);
1036 void gss_sec_gc_ctx_kr(struct ptlrpc_sec *sec)
1038 struct gss_sec_keyring *gsec_kr = sec2gsec_keyring(sec);
1039 struct hlist_head freelist = HLIST_HEAD_INIT;
1040 struct hlist_node __maybe_unused *pos, *next;
1041 struct ptlrpc_cli_ctx *ctx;
1044 CWARN("running gc\n");
1046 spin_lock(&sec->ps_lock);
1047 cfs_hlist_for_each_entry_safe(ctx, pos, next,
1048 &gsec_kr->gsk_clist, cc_cache) {
1049 LASSERT(atomic_read(&ctx->cc_refcount) > 0);
1051 atomic_inc(&ctx->cc_refcount);
1053 if (cli_ctx_check_death(ctx) && ctx_unlist_kr(ctx, 1)) {
1054 hlist_add_head(&ctx->cc_cache, &freelist);
1055 CWARN("unhashed ctx %p\n", ctx);
1057 LASSERT(atomic_read(&ctx->cc_refcount) >= 2);
1058 atomic_dec(&ctx->cc_refcount);
1061 spin_unlock(&sec->ps_lock);
1063 dispose_ctx_list_kr(&freelist);
1069 int gss_sec_display_kr(struct ptlrpc_sec *sec, struct seq_file *seq)
1071 struct gss_sec_keyring *gsec_kr = sec2gsec_keyring(sec);
1072 struct hlist_node __maybe_unused *pos, *next;
1073 struct ptlrpc_cli_ctx *ctx;
1074 struct gss_cli_ctx *gctx;
1075 time64_t now = ktime_get_real_seconds();
1078 spin_lock(&sec->ps_lock);
1079 cfs_hlist_for_each_entry_safe(ctx, pos, next,
1080 &gsec_kr->gsk_clist, cc_cache) {
1085 gctx = ctx2gctx(ctx);
1086 key = ctx2gctx_keyring(ctx)->gck_key;
1088 gss_cli_ctx_flags2str(ctx->cc_flags,
1089 flags_str, sizeof(flags_str));
1091 if (gctx->gc_mechctx)
1092 lgss_display(gctx->gc_mechctx, mech, sizeof(mech));
1094 snprintf(mech, sizeof(mech), "N/A");
1095 mech[sizeof(mech) - 1] = '\0';
1098 "%p: uid %u, ref %d, expire %lld(%+lld), fl %s, seq %d, win %u, key %08x(ref %d), hdl %#llx:%#llx, mech: %s\n",
1099 ctx, ctx->cc_vcred.vc_uid,
1100 atomic_read(&ctx->cc_refcount),
1102 ctx->cc_expire ? ctx->cc_expire - now : 0,
1104 atomic_read(&gctx->gc_seq),
1106 key ? key->serial : 0,
1107 key ? ll_read_key_usage(key) : 0,
1108 gss_handle_to_u64(&gctx->gc_handle),
1109 gss_handle_to_u64(&gctx->gc_svc_handle),
1112 spin_unlock(&sec->ps_lock);
1117 /****************************************
1119 ****************************************/
1122 int gss_cli_ctx_refresh_kr(struct ptlrpc_cli_ctx *ctx)
1124 /* upcall is already on the way */
1129 int gss_cli_ctx_validate_kr(struct ptlrpc_cli_ctx *ctx)
1131 LASSERT(atomic_read(&ctx->cc_refcount) > 0);
1132 LASSERT(ctx->cc_sec);
1134 if (cli_ctx_check_death(ctx)) {
1139 if (cli_ctx_is_ready(ctx))
1145 void gss_cli_ctx_die_kr(struct ptlrpc_cli_ctx *ctx, int grace)
1147 LASSERT(atomic_read(&ctx->cc_refcount) > 0);
1148 LASSERT(ctx->cc_sec);
1150 cli_ctx_expire(ctx);
1154 /****************************************
1155 * (reverse) service *
1156 ****************************************/
1159 * reverse context could have nothing to do with keyrings. here we still keep
1160 * the version which bind to a key, for future reference.
1162 #define HAVE_REVERSE_CTX_NOKEY
1164 #ifdef HAVE_REVERSE_CTX_NOKEY
1167 int sec_install_rctx_kr(struct ptlrpc_sec *sec,
1168 struct ptlrpc_svc_ctx *svc_ctx)
1170 struct ptlrpc_cli_ctx *cli_ctx;
1171 struct vfs_cred vcred = { .vc_uid = 0 };
1177 cli_ctx = ctx_create_kr(sec, &vcred);
1178 if (cli_ctx == NULL)
1181 rc = gss_copy_rvc_cli_ctx(cli_ctx, svc_ctx);
1183 CERROR("failed copy reverse cli ctx: %d\n", rc);
1185 ctx_put_kr(cli_ctx, 1);
1189 rvs_sec_install_root_ctx_kr(sec, cli_ctx, NULL);
1191 ctx_put_kr(cli_ctx, 1);
1196 #else /* ! HAVE_REVERSE_CTX_NOKEY */
1199 int sec_install_rctx_kr(struct ptlrpc_sec *sec,
1200 struct ptlrpc_svc_ctx *svc_ctx)
1202 struct ptlrpc_cli_ctx *cli_ctx = NULL;
1204 struct vfs_cred vcred = { .vc_uid = 0 };
1212 construct_key_desc(desc, sizeof(desc), sec, 0);
1214 key = key_alloc(&gss_key_type, desc, 0, 0,
1215 KEY_POS_ALL | KEY_USR_ALL, 1);
1217 CERROR("failed to alloc key: %ld\n", PTR_ERR(key));
1218 return PTR_ERR(key);
1221 rc = key_instantiate_and_link(key, NULL, 0, NULL, NULL);
1223 CERROR("failed to instantiate key: %d\n", rc);
1227 down_write(&key->sem);
1229 LASSERT(!key_get_payload(key, 0));
1231 cli_ctx = ctx_create_kr(sec, &vcred);
1232 if (cli_ctx == NULL) {
1237 rc = gss_copy_rvc_cli_ctx(cli_ctx, svc_ctx);
1239 CERROR("failed copy reverse cli ctx: %d\n", rc);
1243 rvs_sec_install_root_ctx_kr(sec, cli_ctx, key);
1245 ctx_put_kr(cli_ctx, 1);
1246 up_write(&key->sem);
1255 ctx_put_kr(cli_ctx, 1);
1257 up_write(&key->sem);
1263 #endif /* HAVE_REVERSE_CTX_NOKEY */
1265 /****************************************
1267 ****************************************/
1270 int gss_svc_accept_kr(struct ptlrpc_request *req)
1272 return gss_svc_accept(&gss_policy_keyring, req);
1276 int gss_svc_install_rctx_kr(struct obd_import *imp,
1277 struct ptlrpc_svc_ctx *svc_ctx)
1279 struct ptlrpc_sec *sec;
1282 sec = sptlrpc_import_sec_ref(imp);
1285 rc = sec_install_rctx_kr(sec, svc_ctx);
1286 sptlrpc_sec_put(sec);
1291 /****************************************
1293 ****************************************/
1296 #ifdef HAVE_KEY_TYPE_INSTANTIATE_2ARGS
1297 int gss_kt_instantiate(struct key *key, struct key_preparsed_payload *prep)
1299 const void *data = prep->data;
1300 size_t datalen = prep->datalen;
1302 int gss_kt_instantiate(struct key *key, const void *data, size_t datalen)
1308 if (data != NULL || datalen != 0) {
1309 CERROR("invalid: data %p, len %lu\n", data, (long)datalen);
1313 if (key_get_payload(key, 0)) {
1314 CERROR("key already have payload\n");
1318 /* link the key to session keyring, so following context negotiation
1319 * rpc fired from user space could find this key. This will be unlinked
1320 * automatically when upcall processes die.
1322 * we can't do this through keyctl from userspace, because the upcall
1323 * might be neither possessor nor owner of the key (setuid).
1325 * the session keyring is created upon upcall, and don't change all
1326 * the way until upcall finished, so rcu lock is not needed here.
1328 LASSERT(key_tgcred(current)->session_keyring);
1331 rc = key_link(key_tgcred(current)->session_keyring, key);
1334 CERROR("failed to link key %08x to keyring %08x: %d\n",
1336 key_tgcred(current)->session_keyring->serial, rc);
1340 CDEBUG(D_SEC, "key %p instantiated, ctx %p\n", key,
1341 key_get_payload(key, 0));
1346 * called with key semaphore write locked. it means we can operate
1347 * on the context without fear of loosing refcount.
1350 #ifdef HAVE_KEY_TYPE_INSTANTIATE_2ARGS
1351 int gss_kt_update(struct key *key, struct key_preparsed_payload *prep)
1353 const void *data = prep->data;
1354 __u32 datalen32 = (__u32) prep->datalen;
1356 int gss_kt_update(struct key *key, const void *data, size_t datalen)
1358 __u32 datalen32 = (__u32) datalen;
1360 struct ptlrpc_cli_ctx *ctx = key_get_payload(key, 0);
1361 struct gss_cli_ctx *gctx;
1362 rawobj_t tmpobj = RAWOBJ_EMPTY;
1366 if (data == NULL || datalen32 == 0) {
1367 CWARN("invalid: data %p, len %lu\n", data, (long)datalen32);
1371 /* if upcall finished negotiation too fast (mostly likely because
1372 * of local error happened) and call kt_update(), the ctx
1373 * might be still NULL. but the key will finally be associate
1374 * with a context, or be revoked. if key status is fine, return
1375 * -EAGAIN to allow userspace sleep a while and call again. */
1377 CDEBUG(D_SEC, "update too soon: key %p(%x) flags %lx\n",
1378 key, key->serial, key->flags);
1380 rc = key_validate(key);
1387 LASSERT(atomic_read(&ctx->cc_refcount) > 0);
1388 LASSERT(ctx->cc_sec);
1390 ctx_clear_timer_kr(ctx);
1392 /* don't proceed if already refreshed */
1393 if (cli_ctx_is_refreshed(ctx)) {
1394 CWARN("ctx already done refresh\n");
1398 sptlrpc_cli_ctx_get(ctx);
1399 gctx = ctx2gctx(ctx);
1401 rc = buffer_extract_bytes(&data, &datalen32, &gctx->gc_win,
1402 sizeof(gctx->gc_win));
1404 CERROR("failed extract seq_win\n");
1408 if (gctx->gc_win == 0) {
1409 __u32 nego_rpc_err, nego_gss_err;
1411 rc = buffer_extract_bytes(&data, &datalen32, &nego_rpc_err,
1412 sizeof(nego_rpc_err));
1414 CERROR("cannot extract RPC: rc = %d\n", rc);
1418 rc = buffer_extract_bytes(&data, &datalen32, &nego_gss_err,
1419 sizeof(nego_gss_err));
1421 CERROR("failed to extract gss rc = %d\n", rc);
1425 CERROR("negotiation: rpc err %d, gss err %x\n",
1426 nego_rpc_err, nego_gss_err);
1428 rc = nego_rpc_err ? nego_rpc_err : -EACCES;
1430 rc = rawobj_extract_local_alloc(&gctx->gc_handle,
1431 (__u32 **) &data, &datalen32);
1433 CERROR("failed extract handle\n");
1437 rc = rawobj_extract_local(&tmpobj,
1438 (__u32 **) &data, &datalen32);
1440 CERROR("failed extract mech\n");
1444 rc = lgss_import_sec_context(&tmpobj,
1445 sec2gsec(ctx->cc_sec)->gs_mech,
1447 if (rc != GSS_S_COMPLETE)
1448 CERROR("failed import context\n");
1453 /* we don't care what current status of this ctx, even someone else
1454 * is operating on the ctx at the same time. we just add up our own
1457 gss_cli_ctx_uptodate(gctx);
1459 /* this will also revoke the key. has to be done before
1460 * wakeup waiters otherwise they can find the stale key */
1461 kill_key_locked(key);
1463 cli_ctx_expire(ctx);
1465 if (rc != -ERESTART)
1466 set_bit(PTLRPC_CTX_ERROR_BIT, &ctx->cc_flags);
1469 /* let user space think it's a success */
1470 sptlrpc_cli_ctx_put(ctx, 1);
1474 #ifndef HAVE_KEY_MATCH_DATA
1476 gss_kt_match(const struct key *key, const void *desc)
1478 return strcmp(key->description, (const char *) desc) == 0 &&
1479 !test_bit(KEY_FLAG_REVOKED, &key->flags);
1481 #else /* ! HAVE_KEY_MATCH_DATA */
1483 gss_kt_match(const struct key *key, const struct key_match_data *match_data)
1485 const char *desc = match_data->raw_data;
1487 return strcmp(key->description, desc) == 0 &&
1488 !test_bit(KEY_FLAG_REVOKED, &key->flags);
1492 * Preparse the match criterion.
1494 static int gss_kt_match_preparse(struct key_match_data *match_data)
1496 match_data->lookup_type = KEYRING_SEARCH_LOOKUP_DIRECT;
1497 match_data->cmp = gss_kt_match;
1500 #endif /* HAVE_KEY_MATCH_DATA */
1503 void gss_kt_destroy(struct key *key)
1506 LASSERT(!key_get_payload(key, 0));
1507 CDEBUG(D_SEC, "destroy key %p\n", key);
1512 void gss_kt_describe(const struct key *key, struct seq_file *s)
1514 if (key->description == NULL)
1515 seq_puts(s, "[null]");
1517 seq_puts(s, key->description);
1520 static struct key_type gss_key_type =
1524 .instantiate = gss_kt_instantiate,
1525 .update = gss_kt_update,
1526 #ifdef HAVE_KEY_MATCH_DATA
1527 .match_preparse = gss_kt_match_preparse,
1529 .match = gss_kt_match,
1531 .destroy = gss_kt_destroy,
1532 .describe = gss_kt_describe,
1535 /****************************************
1536 * lustre gss keyring policy *
1537 ****************************************/
1539 static struct ptlrpc_ctx_ops gss_keyring_ctxops = {
1540 .match = gss_cli_ctx_match,
1541 .refresh = gss_cli_ctx_refresh_kr,
1542 .validate = gss_cli_ctx_validate_kr,
1543 .die = gss_cli_ctx_die_kr,
1544 .sign = gss_cli_ctx_sign,
1545 .verify = gss_cli_ctx_verify,
1546 .seal = gss_cli_ctx_seal,
1547 .unseal = gss_cli_ctx_unseal,
1548 .wrap_bulk = gss_cli_ctx_wrap_bulk,
1549 .unwrap_bulk = gss_cli_ctx_unwrap_bulk,
1552 static struct ptlrpc_sec_cops gss_sec_keyring_cops = {
1553 .create_sec = gss_sec_create_kr,
1554 .destroy_sec = gss_sec_destroy_kr,
1555 .kill_sec = gss_sec_kill,
1556 .lookup_ctx = gss_sec_lookup_ctx_kr,
1557 .release_ctx = gss_sec_release_ctx_kr,
1558 .flush_ctx_cache = gss_sec_flush_ctx_cache_kr,
1559 .gc_ctx = gss_sec_gc_ctx_kr,
1560 .install_rctx = gss_sec_install_rctx,
1561 .alloc_reqbuf = gss_alloc_reqbuf,
1562 .free_reqbuf = gss_free_reqbuf,
1563 .alloc_repbuf = gss_alloc_repbuf,
1564 .free_repbuf = gss_free_repbuf,
1565 .enlarge_reqbuf = gss_enlarge_reqbuf,
1566 .display = gss_sec_display_kr,
1569 static struct ptlrpc_sec_sops gss_sec_keyring_sops = {
1570 .accept = gss_svc_accept_kr,
1571 .invalidate_ctx = gss_svc_invalidate_ctx,
1572 .alloc_rs = gss_svc_alloc_rs,
1573 .authorize = gss_svc_authorize,
1574 .free_rs = gss_svc_free_rs,
1575 .free_ctx = gss_svc_free_ctx,
1576 .prep_bulk = gss_svc_prep_bulk,
1577 .unwrap_bulk = gss_svc_unwrap_bulk,
1578 .wrap_bulk = gss_svc_wrap_bulk,
1579 .install_rctx = gss_svc_install_rctx_kr,
1582 static struct ptlrpc_sec_policy gss_policy_keyring = {
1583 .sp_owner = THIS_MODULE,
1584 .sp_name = "gss.keyring",
1585 .sp_policy = SPTLRPC_POLICY_GSS,
1586 .sp_cops = &gss_sec_keyring_cops,
1587 .sp_sops = &gss_sec_keyring_sops,
1591 int __init gss_init_keyring(void)
1595 rc = register_key_type(&gss_key_type);
1597 CERROR("failed to register keyring type: %d\n", rc);
1601 rc = sptlrpc_register_policy(&gss_policy_keyring);
1603 unregister_key_type(&gss_key_type);
1610 void __exit gss_exit_keyring(void)
1612 unregister_key_type(&gss_key_type);
1613 sptlrpc_unregister_policy(&gss_policy_keyring);