Loading...
1#include <linux/bpf.h>
2#include <linux/errno.h>
3#include <linux/errqueue.h>
4#include <linux/file.h>
5#include <linux/in.h>
6#include <linux/kernel.h>
7#include <linux/module.h>
8#include <linux/net.h>
9#include <linux/netdevice.h>
10#include <linux/poll.h>
11#include <linux/rculist.h>
12#include <linux/skbuff.h>
13#include <linux/socket.h>
14#include <linux/uaccess.h>
15#include <linux/workqueue.h>
16#include <net/kcm.h>
17#include <net/netns/generic.h>
18#include <net/sock.h>
19#include <net/tcp.h>
20#include <uapi/linux/kcm.h>
21
22unsigned int kcm_net_id;
23
24static struct kmem_cache *kcm_psockp __read_mostly;
25static struct kmem_cache *kcm_muxp __read_mostly;
26static struct workqueue_struct *kcm_wq;
27
28static inline struct kcm_sock *kcm_sk(const struct sock *sk)
29{
30 return (struct kcm_sock *)sk;
31}
32
33static inline struct kcm_tx_msg *kcm_tx_msg(struct sk_buff *skb)
34{
35 return (struct kcm_tx_msg *)skb->cb;
36}
37
38static inline struct kcm_rx_msg *kcm_rx_msg(struct sk_buff *skb)
39{
40 return (struct kcm_rx_msg *)((void *)skb->cb +
41 offsetof(struct qdisc_skb_cb, data));
42}
43
44static void report_csk_error(struct sock *csk, int err)
45{
46 csk->sk_err = EPIPE;
47 csk->sk_error_report(csk);
48}
49
50/* Callback lock held */
51static void kcm_abort_rx_psock(struct kcm_psock *psock, int err,
52 struct sk_buff *skb)
53{
54 struct sock *csk = psock->sk;
55
56 /* Unrecoverable error in receive */
57
58 del_timer(&psock->rx_msg_timer);
59
60 if (psock->rx_stopped)
61 return;
62
63 psock->rx_stopped = 1;
64 KCM_STATS_INCR(psock->stats.rx_aborts);
65
66 /* Report an error on the lower socket */
67 report_csk_error(csk, err);
68}
69
70static void kcm_abort_tx_psock(struct kcm_psock *psock, int err,
71 bool wakeup_kcm)
72{
73 struct sock *csk = psock->sk;
74 struct kcm_mux *mux = psock->mux;
75
76 /* Unrecoverable error in transmit */
77
78 spin_lock_bh(&mux->lock);
79
80 if (psock->tx_stopped) {
81 spin_unlock_bh(&mux->lock);
82 return;
83 }
84
85 psock->tx_stopped = 1;
86 KCM_STATS_INCR(psock->stats.tx_aborts);
87
88 if (!psock->tx_kcm) {
89 /* Take off psocks_avail list */
90 list_del(&psock->psock_avail_list);
91 } else if (wakeup_kcm) {
92 /* In this case psock is being aborted while outside of
93 * write_msgs and psock is reserved. Schedule tx_work
94 * to handle the failure there. Need to commit tx_stopped
95 * before queuing work.
96 */
97 smp_mb();
98
99 queue_work(kcm_wq, &psock->tx_kcm->tx_work);
100 }
101
102 spin_unlock_bh(&mux->lock);
103
104 /* Report error on lower socket */
105 report_csk_error(csk, err);
106}
107
108/* RX mux lock held. */
109static void kcm_update_rx_mux_stats(struct kcm_mux *mux,
110 struct kcm_psock *psock)
111{
112 KCM_STATS_ADD(mux->stats.rx_bytes,
113 psock->stats.rx_bytes - psock->saved_rx_bytes);
114 mux->stats.rx_msgs +=
115 psock->stats.rx_msgs - psock->saved_rx_msgs;
116 psock->saved_rx_msgs = psock->stats.rx_msgs;
117 psock->saved_rx_bytes = psock->stats.rx_bytes;
118}
119
120static void kcm_update_tx_mux_stats(struct kcm_mux *mux,
121 struct kcm_psock *psock)
122{
123 KCM_STATS_ADD(mux->stats.tx_bytes,
124 psock->stats.tx_bytes - psock->saved_tx_bytes);
125 mux->stats.tx_msgs +=
126 psock->stats.tx_msgs - psock->saved_tx_msgs;
127 psock->saved_tx_msgs = psock->stats.tx_msgs;
128 psock->saved_tx_bytes = psock->stats.tx_bytes;
129}
130
131static int kcm_queue_rcv_skb(struct sock *sk, struct sk_buff *skb);
132
133/* KCM is ready to receive messages on its queue-- either the KCM is new or
134 * has become unblocked after being blocked on full socket buffer. Queue any
135 * pending ready messages on a psock. RX mux lock held.
136 */
137static void kcm_rcv_ready(struct kcm_sock *kcm)
138{
139 struct kcm_mux *mux = kcm->mux;
140 struct kcm_psock *psock;
141 struct sk_buff *skb;
142
143 if (unlikely(kcm->rx_wait || kcm->rx_psock || kcm->rx_disabled))
144 return;
145
146 while (unlikely((skb = __skb_dequeue(&mux->rx_hold_queue)))) {
147 if (kcm_queue_rcv_skb(&kcm->sk, skb)) {
148 /* Assuming buffer limit has been reached */
149 skb_queue_head(&mux->rx_hold_queue, skb);
150 WARN_ON(!sk_rmem_alloc_get(&kcm->sk));
151 return;
152 }
153 }
154
155 while (!list_empty(&mux->psocks_ready)) {
156 psock = list_first_entry(&mux->psocks_ready, struct kcm_psock,
157 psock_ready_list);
158
159 if (kcm_queue_rcv_skb(&kcm->sk, psock->ready_rx_msg)) {
160 /* Assuming buffer limit has been reached */
161 WARN_ON(!sk_rmem_alloc_get(&kcm->sk));
162 return;
163 }
164
165 /* Consumed the ready message on the psock. Schedule rx_work to
166 * get more messages.
167 */
168 list_del(&psock->psock_ready_list);
169 psock->ready_rx_msg = NULL;
170
171 /* Commit clearing of ready_rx_msg for queuing work */
172 smp_mb();
173
174 queue_work(kcm_wq, &psock->rx_work);
175 }
176
177 /* Buffer limit is okay now, add to ready list */
178 list_add_tail(&kcm->wait_rx_list,
179 &kcm->mux->kcm_rx_waiters);
180 kcm->rx_wait = true;
181}
182
183static void kcm_rfree(struct sk_buff *skb)
184{
185 struct sock *sk = skb->sk;
186 struct kcm_sock *kcm = kcm_sk(sk);
187 struct kcm_mux *mux = kcm->mux;
188 unsigned int len = skb->truesize;
189
190 sk_mem_uncharge(sk, len);
191 atomic_sub(len, &sk->sk_rmem_alloc);
192
193 /* For reading rx_wait and rx_psock without holding lock */
194 smp_mb__after_atomic();
195
196 if (!kcm->rx_wait && !kcm->rx_psock &&
197 sk_rmem_alloc_get(sk) < sk->sk_rcvlowat) {
198 spin_lock_bh(&mux->rx_lock);
199 kcm_rcv_ready(kcm);
200 spin_unlock_bh(&mux->rx_lock);
201 }
202}
203
204static int kcm_queue_rcv_skb(struct sock *sk, struct sk_buff *skb)
205{
206 struct sk_buff_head *list = &sk->sk_receive_queue;
207
208 if (atomic_read(&sk->sk_rmem_alloc) >= sk->sk_rcvbuf)
209 return -ENOMEM;
210
211 if (!sk_rmem_schedule(sk, skb, skb->truesize))
212 return -ENOBUFS;
213
214 skb->dev = NULL;
215
216 skb_orphan(skb);
217 skb->sk = sk;
218 skb->destructor = kcm_rfree;
219 atomic_add(skb->truesize, &sk->sk_rmem_alloc);
220 sk_mem_charge(sk, skb->truesize);
221
222 skb_queue_tail(list, skb);
223
224 if (!sock_flag(sk, SOCK_DEAD))
225 sk->sk_data_ready(sk);
226
227 return 0;
228}
229
230/* Requeue received messages for a kcm socket to other kcm sockets. This is
231 * called with a kcm socket is receive disabled.
232 * RX mux lock held.
233 */
234static void requeue_rx_msgs(struct kcm_mux *mux, struct sk_buff_head *head)
235{
236 struct sk_buff *skb;
237 struct kcm_sock *kcm;
238
239 while ((skb = __skb_dequeue(head))) {
240 /* Reset destructor to avoid calling kcm_rcv_ready */
241 skb->destructor = sock_rfree;
242 skb_orphan(skb);
243try_again:
244 if (list_empty(&mux->kcm_rx_waiters)) {
245 skb_queue_tail(&mux->rx_hold_queue, skb);
246 continue;
247 }
248
249 kcm = list_first_entry(&mux->kcm_rx_waiters,
250 struct kcm_sock, wait_rx_list);
251
252 if (kcm_queue_rcv_skb(&kcm->sk, skb)) {
253 /* Should mean socket buffer full */
254 list_del(&kcm->wait_rx_list);
255 kcm->rx_wait = false;
256
257 /* Commit rx_wait to read in kcm_free */
258 smp_wmb();
259
260 goto try_again;
261 }
262 }
263}
264
265/* Lower sock lock held */
266static struct kcm_sock *reserve_rx_kcm(struct kcm_psock *psock,
267 struct sk_buff *head)
268{
269 struct kcm_mux *mux = psock->mux;
270 struct kcm_sock *kcm;
271
272 WARN_ON(psock->ready_rx_msg);
273
274 if (psock->rx_kcm)
275 return psock->rx_kcm;
276
277 spin_lock_bh(&mux->rx_lock);
278
279 if (psock->rx_kcm) {
280 spin_unlock_bh(&mux->rx_lock);
281 return psock->rx_kcm;
282 }
283
284 kcm_update_rx_mux_stats(mux, psock);
285
286 if (list_empty(&mux->kcm_rx_waiters)) {
287 psock->ready_rx_msg = head;
288 list_add_tail(&psock->psock_ready_list,
289 &mux->psocks_ready);
290 spin_unlock_bh(&mux->rx_lock);
291 return NULL;
292 }
293
294 kcm = list_first_entry(&mux->kcm_rx_waiters,
295 struct kcm_sock, wait_rx_list);
296 list_del(&kcm->wait_rx_list);
297 kcm->rx_wait = false;
298
299 psock->rx_kcm = kcm;
300 kcm->rx_psock = psock;
301
302 spin_unlock_bh(&mux->rx_lock);
303
304 return kcm;
305}
306
307static void kcm_done(struct kcm_sock *kcm);
308
309static void kcm_done_work(struct work_struct *w)
310{
311 kcm_done(container_of(w, struct kcm_sock, done_work));
312}
313
314/* Lower sock held */
315static void unreserve_rx_kcm(struct kcm_psock *psock,
316 bool rcv_ready)
317{
318 struct kcm_sock *kcm = psock->rx_kcm;
319 struct kcm_mux *mux = psock->mux;
320
321 if (!kcm)
322 return;
323
324 spin_lock_bh(&mux->rx_lock);
325
326 psock->rx_kcm = NULL;
327 kcm->rx_psock = NULL;
328
329 /* Commit kcm->rx_psock before sk_rmem_alloc_get to sync with
330 * kcm_rfree
331 */
332 smp_mb();
333
334 if (unlikely(kcm->done)) {
335 spin_unlock_bh(&mux->rx_lock);
336
337 /* Need to run kcm_done in a task since we need to qcquire
338 * callback locks which may already be held here.
339 */
340 INIT_WORK(&kcm->done_work, kcm_done_work);
341 schedule_work(&kcm->done_work);
342 return;
343 }
344
345 if (unlikely(kcm->rx_disabled)) {
346 requeue_rx_msgs(mux, &kcm->sk.sk_receive_queue);
347 } else if (rcv_ready || unlikely(!sk_rmem_alloc_get(&kcm->sk))) {
348 /* Check for degenerative race with rx_wait that all
349 * data was dequeued (accounted for in kcm_rfree).
350 */
351 kcm_rcv_ready(kcm);
352 }
353 spin_unlock_bh(&mux->rx_lock);
354}
355
356static void kcm_start_rx_timer(struct kcm_psock *psock)
357{
358 if (psock->sk->sk_rcvtimeo)
359 mod_timer(&psock->rx_msg_timer, psock->sk->sk_rcvtimeo);
360}
361
362/* Macro to invoke filter function. */
363#define KCM_RUN_FILTER(prog, ctx) \
364 (*prog->bpf_func)(ctx, prog->insnsi)
365
366/* Lower socket lock held */
367static int kcm_tcp_recv(read_descriptor_t *desc, struct sk_buff *orig_skb,
368 unsigned int orig_offset, size_t orig_len)
369{
370 struct kcm_psock *psock = (struct kcm_psock *)desc->arg.data;
371 struct kcm_rx_msg *rxm;
372 struct kcm_sock *kcm;
373 struct sk_buff *head, *skb;
374 size_t eaten = 0, cand_len;
375 ssize_t extra;
376 int err;
377 bool cloned_orig = false;
378
379 if (psock->ready_rx_msg)
380 return 0;
381
382 head = psock->rx_skb_head;
383 if (head) {
384 /* Message already in progress */
385
386 rxm = kcm_rx_msg(head);
387 if (unlikely(rxm->early_eaten)) {
388 /* Already some number of bytes on the receive sock
389 * data saved in rx_skb_head, just indicate they
390 * are consumed.
391 */
392 eaten = orig_len <= rxm->early_eaten ?
393 orig_len : rxm->early_eaten;
394 rxm->early_eaten -= eaten;
395
396 return eaten;
397 }
398
399 if (unlikely(orig_offset)) {
400 /* Getting data with a non-zero offset when a message is
401 * in progress is not expected. If it does happen, we
402 * need to clone and pull since we can't deal with
403 * offsets in the skbs for a message expect in the head.
404 */
405 orig_skb = skb_clone(orig_skb, GFP_ATOMIC);
406 if (!orig_skb) {
407 KCM_STATS_INCR(psock->stats.rx_mem_fail);
408 desc->error = -ENOMEM;
409 return 0;
410 }
411 if (!pskb_pull(orig_skb, orig_offset)) {
412 KCM_STATS_INCR(psock->stats.rx_mem_fail);
413 kfree_skb(orig_skb);
414 desc->error = -ENOMEM;
415 return 0;
416 }
417 cloned_orig = true;
418 orig_offset = 0;
419 }
420
421 if (!psock->rx_skb_nextp) {
422 /* We are going to append to the frags_list of head.
423 * Need to unshare the frag_list.
424 */
425 err = skb_unclone(head, GFP_ATOMIC);
426 if (err) {
427 KCM_STATS_INCR(psock->stats.rx_mem_fail);
428 desc->error = err;
429 return 0;
430 }
431
432 if (unlikely(skb_shinfo(head)->frag_list)) {
433 /* We can't append to an sk_buff that already
434 * has a frag_list. We create a new head, point
435 * the frag_list of that to the old head, and
436 * then are able to use the old head->next for
437 * appending to the message.
438 */
439 if (WARN_ON(head->next)) {
440 desc->error = -EINVAL;
441 return 0;
442 }
443
444 skb = alloc_skb(0, GFP_ATOMIC);
445 if (!skb) {
446 KCM_STATS_INCR(psock->stats.rx_mem_fail);
447 desc->error = -ENOMEM;
448 return 0;
449 }
450 skb->len = head->len;
451 skb->data_len = head->len;
452 skb->truesize = head->truesize;
453 *kcm_rx_msg(skb) = *kcm_rx_msg(head);
454 psock->rx_skb_nextp = &head->next;
455 skb_shinfo(skb)->frag_list = head;
456 psock->rx_skb_head = skb;
457 head = skb;
458 } else {
459 psock->rx_skb_nextp =
460 &skb_shinfo(head)->frag_list;
461 }
462 }
463 }
464
465 while (eaten < orig_len) {
466 /* Always clone since we will consume something */
467 skb = skb_clone(orig_skb, GFP_ATOMIC);
468 if (!skb) {
469 KCM_STATS_INCR(psock->stats.rx_mem_fail);
470 desc->error = -ENOMEM;
471 break;
472 }
473
474 cand_len = orig_len - eaten;
475
476 head = psock->rx_skb_head;
477 if (!head) {
478 head = skb;
479 psock->rx_skb_head = head;
480 /* Will set rx_skb_nextp on next packet if needed */
481 psock->rx_skb_nextp = NULL;
482 rxm = kcm_rx_msg(head);
483 memset(rxm, 0, sizeof(*rxm));
484 rxm->offset = orig_offset + eaten;
485 } else {
486 /* Unclone since we may be appending to an skb that we
487 * already share a frag_list with.
488 */
489 err = skb_unclone(skb, GFP_ATOMIC);
490 if (err) {
491 KCM_STATS_INCR(psock->stats.rx_mem_fail);
492 desc->error = err;
493 break;
494 }
495
496 rxm = kcm_rx_msg(head);
497 *psock->rx_skb_nextp = skb;
498 psock->rx_skb_nextp = &skb->next;
499 head->data_len += skb->len;
500 head->len += skb->len;
501 head->truesize += skb->truesize;
502 }
503
504 if (!rxm->full_len) {
505 ssize_t len;
506
507 len = KCM_RUN_FILTER(psock->bpf_prog, head);
508
509 if (!len) {
510 /* Need more header to determine length */
511 if (!rxm->accum_len) {
512 /* Start RX timer for new message */
513 kcm_start_rx_timer(psock);
514 }
515 rxm->accum_len += cand_len;
516 eaten += cand_len;
517 KCM_STATS_INCR(psock->stats.rx_need_more_hdr);
518 WARN_ON(eaten != orig_len);
519 break;
520 } else if (len > psock->sk->sk_rcvbuf) {
521 /* Message length exceeds maximum allowed */
522 KCM_STATS_INCR(psock->stats.rx_msg_too_big);
523 desc->error = -EMSGSIZE;
524 psock->rx_skb_head = NULL;
525 kcm_abort_rx_psock(psock, EMSGSIZE, head);
526 break;
527 } else if (len <= (ssize_t)head->len -
528 skb->len - rxm->offset) {
529 /* Length must be into new skb (and also
530 * greater than zero)
531 */
532 KCM_STATS_INCR(psock->stats.rx_bad_hdr_len);
533 desc->error = -EPROTO;
534 psock->rx_skb_head = NULL;
535 kcm_abort_rx_psock(psock, EPROTO, head);
536 break;
537 }
538
539 rxm->full_len = len;
540 }
541
542 extra = (ssize_t)(rxm->accum_len + cand_len) - rxm->full_len;
543
544 if (extra < 0) {
545 /* Message not complete yet. */
546 if (rxm->full_len - rxm->accum_len >
547 tcp_inq(psock->sk)) {
548 /* Don't have the whole messages in the socket
549 * buffer. Set psock->rx_need_bytes to wait for
550 * the rest of the message. Also, set "early
551 * eaten" since we've already buffered the skb
552 * but don't consume yet per tcp_read_sock.
553 */
554
555 if (!rxm->accum_len) {
556 /* Start RX timer for new message */
557 kcm_start_rx_timer(psock);
558 }
559
560 psock->rx_need_bytes = rxm->full_len -
561 rxm->accum_len;
562 rxm->accum_len += cand_len;
563 rxm->early_eaten = cand_len;
564 KCM_STATS_ADD(psock->stats.rx_bytes, cand_len);
565 desc->count = 0; /* Stop reading socket */
566 break;
567 }
568 rxm->accum_len += cand_len;
569 eaten += cand_len;
570 WARN_ON(eaten != orig_len);
571 break;
572 }
573
574 /* Positive extra indicates ore bytes than needed for the
575 * message
576 */
577
578 WARN_ON(extra > cand_len);
579
580 eaten += (cand_len - extra);
581
582 /* Hurray, we have a new message! */
583 del_timer(&psock->rx_msg_timer);
584 psock->rx_skb_head = NULL;
585 KCM_STATS_INCR(psock->stats.rx_msgs);
586
587try_queue:
588 kcm = reserve_rx_kcm(psock, head);
589 if (!kcm) {
590 /* Unable to reserve a KCM, message is held in psock. */
591 break;
592 }
593
594 if (kcm_queue_rcv_skb(&kcm->sk, head)) {
595 /* Should mean socket buffer full */
596 unreserve_rx_kcm(psock, false);
597 goto try_queue;
598 }
599 }
600
601 if (cloned_orig)
602 kfree_skb(orig_skb);
603
604 KCM_STATS_ADD(psock->stats.rx_bytes, eaten);
605
606 return eaten;
607}
608
609/* Called with lock held on lower socket */
610static int psock_tcp_read_sock(struct kcm_psock *psock)
611{
612 read_descriptor_t desc;
613
614 desc.arg.data = psock;
615 desc.error = 0;
616 desc.count = 1; /* give more than one skb per call */
617
618 /* sk should be locked here, so okay to do tcp_read_sock */
619 tcp_read_sock(psock->sk, &desc, kcm_tcp_recv);
620
621 unreserve_rx_kcm(psock, true);
622
623 return desc.error;
624}
625
626/* Lower sock lock held */
627static void psock_tcp_data_ready(struct sock *sk)
628{
629 struct kcm_psock *psock;
630
631 read_lock_bh(&sk->sk_callback_lock);
632
633 psock = (struct kcm_psock *)sk->sk_user_data;
634 if (unlikely(!psock || psock->rx_stopped))
635 goto out;
636
637 if (psock->ready_rx_msg)
638 goto out;
639
640 if (psock->rx_need_bytes) {
641 if (tcp_inq(sk) >= psock->rx_need_bytes)
642 psock->rx_need_bytes = 0;
643 else
644 goto out;
645 }
646
647 if (psock_tcp_read_sock(psock) == -ENOMEM)
648 queue_delayed_work(kcm_wq, &psock->rx_delayed_work, 0);
649
650out:
651 read_unlock_bh(&sk->sk_callback_lock);
652}
653
654static void do_psock_rx_work(struct kcm_psock *psock)
655{
656 read_descriptor_t rd_desc;
657 struct sock *csk = psock->sk;
658
659 /* We need the read lock to synchronize with psock_tcp_data_ready. We
660 * need the socket lock for calling tcp_read_sock.
661 */
662 lock_sock(csk);
663 read_lock_bh(&csk->sk_callback_lock);
664
665 if (unlikely(csk->sk_user_data != psock))
666 goto out;
667
668 if (unlikely(psock->rx_stopped))
669 goto out;
670
671 if (psock->ready_rx_msg)
672 goto out;
673
674 rd_desc.arg.data = psock;
675
676 if (psock_tcp_read_sock(psock) == -ENOMEM)
677 queue_delayed_work(kcm_wq, &psock->rx_delayed_work, 0);
678
679out:
680 read_unlock_bh(&csk->sk_callback_lock);
681 release_sock(csk);
682}
683
684static void psock_rx_work(struct work_struct *w)
685{
686 do_psock_rx_work(container_of(w, struct kcm_psock, rx_work));
687}
688
689static void psock_rx_delayed_work(struct work_struct *w)
690{
691 do_psock_rx_work(container_of(w, struct kcm_psock,
692 rx_delayed_work.work));
693}
694
695static void psock_tcp_state_change(struct sock *sk)
696{
697 /* TCP only does a POLLIN for a half close. Do a POLLHUP here
698 * since application will normally not poll with POLLIN
699 * on the TCP sockets.
700 */
701
702 report_csk_error(sk, EPIPE);
703}
704
705static void psock_tcp_write_space(struct sock *sk)
706{
707 struct kcm_psock *psock;
708 struct kcm_mux *mux;
709 struct kcm_sock *kcm;
710
711 read_lock_bh(&sk->sk_callback_lock);
712
713 psock = (struct kcm_psock *)sk->sk_user_data;
714 if (unlikely(!psock))
715 goto out;
716
717 mux = psock->mux;
718
719 spin_lock_bh(&mux->lock);
720
721 /* Check if the socket is reserved so someone is waiting for sending. */
722 kcm = psock->tx_kcm;
723 if (kcm)
724 queue_work(kcm_wq, &kcm->tx_work);
725
726 spin_unlock_bh(&mux->lock);
727out:
728 read_unlock_bh(&sk->sk_callback_lock);
729}
730
731static void unreserve_psock(struct kcm_sock *kcm);
732
733/* kcm sock is locked. */
734static struct kcm_psock *reserve_psock(struct kcm_sock *kcm)
735{
736 struct kcm_mux *mux = kcm->mux;
737 struct kcm_psock *psock;
738
739 psock = kcm->tx_psock;
740
741 smp_rmb(); /* Must read tx_psock before tx_wait */
742
743 if (psock) {
744 WARN_ON(kcm->tx_wait);
745 if (unlikely(psock->tx_stopped))
746 unreserve_psock(kcm);
747 else
748 return kcm->tx_psock;
749 }
750
751 spin_lock_bh(&mux->lock);
752
753 /* Check again under lock to see if psock was reserved for this
754 * psock via psock_unreserve.
755 */
756 psock = kcm->tx_psock;
757 if (unlikely(psock)) {
758 WARN_ON(kcm->tx_wait);
759 spin_unlock_bh(&mux->lock);
760 return kcm->tx_psock;
761 }
762
763 if (!list_empty(&mux->psocks_avail)) {
764 psock = list_first_entry(&mux->psocks_avail,
765 struct kcm_psock,
766 psock_avail_list);
767 list_del(&psock->psock_avail_list);
768 if (kcm->tx_wait) {
769 list_del(&kcm->wait_psock_list);
770 kcm->tx_wait = false;
771 }
772 kcm->tx_psock = psock;
773 psock->tx_kcm = kcm;
774 KCM_STATS_INCR(psock->stats.reserved);
775 } else if (!kcm->tx_wait) {
776 list_add_tail(&kcm->wait_psock_list,
777 &mux->kcm_tx_waiters);
778 kcm->tx_wait = true;
779 }
780
781 spin_unlock_bh(&mux->lock);
782
783 return psock;
784}
785
786/* mux lock held */
787static void psock_now_avail(struct kcm_psock *psock)
788{
789 struct kcm_mux *mux = psock->mux;
790 struct kcm_sock *kcm;
791
792 if (list_empty(&mux->kcm_tx_waiters)) {
793 list_add_tail(&psock->psock_avail_list,
794 &mux->psocks_avail);
795 } else {
796 kcm = list_first_entry(&mux->kcm_tx_waiters,
797 struct kcm_sock,
798 wait_psock_list);
799 list_del(&kcm->wait_psock_list);
800 kcm->tx_wait = false;
801 psock->tx_kcm = kcm;
802
803 /* Commit before changing tx_psock since that is read in
804 * reserve_psock before queuing work.
805 */
806 smp_mb();
807
808 kcm->tx_psock = psock;
809 KCM_STATS_INCR(psock->stats.reserved);
810 queue_work(kcm_wq, &kcm->tx_work);
811 }
812}
813
814/* kcm sock is locked. */
815static void unreserve_psock(struct kcm_sock *kcm)
816{
817 struct kcm_psock *psock;
818 struct kcm_mux *mux = kcm->mux;
819
820 spin_lock_bh(&mux->lock);
821
822 psock = kcm->tx_psock;
823
824 if (WARN_ON(!psock)) {
825 spin_unlock_bh(&mux->lock);
826 return;
827 }
828
829 smp_rmb(); /* Read tx_psock before tx_wait */
830
831 kcm_update_tx_mux_stats(mux, psock);
832
833 WARN_ON(kcm->tx_wait);
834
835 kcm->tx_psock = NULL;
836 psock->tx_kcm = NULL;
837 KCM_STATS_INCR(psock->stats.unreserved);
838
839 if (unlikely(psock->tx_stopped)) {
840 if (psock->done) {
841 /* Deferred free */
842 list_del(&psock->psock_list);
843 mux->psocks_cnt--;
844 sock_put(psock->sk);
845 fput(psock->sk->sk_socket->file);
846 kmem_cache_free(kcm_psockp, psock);
847 }
848
849 /* Don't put back on available list */
850
851 spin_unlock_bh(&mux->lock);
852
853 return;
854 }
855
856 psock_now_avail(psock);
857
858 spin_unlock_bh(&mux->lock);
859}
860
861static void kcm_report_tx_retry(struct kcm_sock *kcm)
862{
863 struct kcm_mux *mux = kcm->mux;
864
865 spin_lock_bh(&mux->lock);
866 KCM_STATS_INCR(mux->stats.tx_retries);
867 spin_unlock_bh(&mux->lock);
868}
869
870/* Write any messages ready on the kcm socket. Called with kcm sock lock
871 * held. Return bytes actually sent or error.
872 */
873static int kcm_write_msgs(struct kcm_sock *kcm)
874{
875 struct sock *sk = &kcm->sk;
876 struct kcm_psock *psock;
877 struct sk_buff *skb, *head;
878 struct kcm_tx_msg *txm;
879 unsigned short fragidx, frag_offset;
880 unsigned int sent, total_sent = 0;
881 int ret = 0;
882
883 kcm->tx_wait_more = false;
884 psock = kcm->tx_psock;
885 if (unlikely(psock && psock->tx_stopped)) {
886 /* A reserved psock was aborted asynchronously. Unreserve
887 * it and we'll retry the message.
888 */
889 unreserve_psock(kcm);
890 kcm_report_tx_retry(kcm);
891 if (skb_queue_empty(&sk->sk_write_queue))
892 return 0;
893
894 kcm_tx_msg(skb_peek(&sk->sk_write_queue))->sent = 0;
895
896 } else if (skb_queue_empty(&sk->sk_write_queue)) {
897 return 0;
898 }
899
900 head = skb_peek(&sk->sk_write_queue);
901 txm = kcm_tx_msg(head);
902
903 if (txm->sent) {
904 /* Send of first skbuff in queue already in progress */
905 if (WARN_ON(!psock)) {
906 ret = -EINVAL;
907 goto out;
908 }
909 sent = txm->sent;
910 frag_offset = txm->frag_offset;
911 fragidx = txm->fragidx;
912 skb = txm->frag_skb;
913
914 goto do_frag;
915 }
916
917try_again:
918 psock = reserve_psock(kcm);
919 if (!psock)
920 goto out;
921
922 do {
923 skb = head;
924 txm = kcm_tx_msg(head);
925 sent = 0;
926
927do_frag_list:
928 if (WARN_ON(!skb_shinfo(skb)->nr_frags)) {
929 ret = -EINVAL;
930 goto out;
931 }
932
933 for (fragidx = 0; fragidx < skb_shinfo(skb)->nr_frags;
934 fragidx++) {
935 skb_frag_t *frag;
936
937 frag_offset = 0;
938do_frag:
939 frag = &skb_shinfo(skb)->frags[fragidx];
940 if (WARN_ON(!frag->size)) {
941 ret = -EINVAL;
942 goto out;
943 }
944
945 ret = kernel_sendpage(psock->sk->sk_socket,
946 frag->page.p,
947 frag->page_offset + frag_offset,
948 frag->size - frag_offset,
949 MSG_DONTWAIT);
950 if (ret <= 0) {
951 if (ret == -EAGAIN) {
952 /* Save state to try again when there's
953 * write space on the socket
954 */
955 txm->sent = sent;
956 txm->frag_offset = frag_offset;
957 txm->fragidx = fragidx;
958 txm->frag_skb = skb;
959
960 ret = 0;
961 goto out;
962 }
963
964 /* Hard failure in sending message, abort this
965 * psock since it has lost framing
966 * synchonization and retry sending the
967 * message from the beginning.
968 */
969 kcm_abort_tx_psock(psock, ret ? -ret : EPIPE,
970 true);
971 unreserve_psock(kcm);
972
973 txm->sent = 0;
974 kcm_report_tx_retry(kcm);
975 ret = 0;
976
977 goto try_again;
978 }
979
980 sent += ret;
981 frag_offset += ret;
982 KCM_STATS_ADD(psock->stats.tx_bytes, ret);
983 if (frag_offset < frag->size) {
984 /* Not finished with this frag */
985 goto do_frag;
986 }
987 }
988
989 if (skb == head) {
990 if (skb_has_frag_list(skb)) {
991 skb = skb_shinfo(skb)->frag_list;
992 goto do_frag_list;
993 }
994 } else if (skb->next) {
995 skb = skb->next;
996 goto do_frag_list;
997 }
998
999 /* Successfully sent the whole packet, account for it. */
1000 skb_dequeue(&sk->sk_write_queue);
1001 kfree_skb(head);
1002 sk->sk_wmem_queued -= sent;
1003 total_sent += sent;
1004 KCM_STATS_INCR(psock->stats.tx_msgs);
1005 } while ((head = skb_peek(&sk->sk_write_queue)));
1006out:
1007 if (!head) {
1008 /* Done with all queued messages. */
1009 WARN_ON(!skb_queue_empty(&sk->sk_write_queue));
1010 unreserve_psock(kcm);
1011 }
1012
1013 /* Check if write space is available */
1014 sk->sk_write_space(sk);
1015
1016 return total_sent ? : ret;
1017}
1018
1019static void kcm_tx_work(struct work_struct *w)
1020{
1021 struct kcm_sock *kcm = container_of(w, struct kcm_sock, tx_work);
1022 struct sock *sk = &kcm->sk;
1023 int err;
1024
1025 lock_sock(sk);
1026
1027 /* Primarily for SOCK_DGRAM sockets, also handle asynchronous tx
1028 * aborts
1029 */
1030 err = kcm_write_msgs(kcm);
1031 if (err < 0) {
1032 /* Hard failure in write, report error on KCM socket */
1033 pr_warn("KCM: Hard failure on kcm_write_msgs %d\n", err);
1034 report_csk_error(&kcm->sk, -err);
1035 goto out;
1036 }
1037
1038 /* Primarily for SOCK_SEQPACKET sockets */
1039 if (likely(sk->sk_socket) &&
1040 test_bit(SOCK_NOSPACE, &sk->sk_socket->flags)) {
1041 clear_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
1042 sk->sk_write_space(sk);
1043 }
1044
1045out:
1046 release_sock(sk);
1047}
1048
1049static void kcm_push(struct kcm_sock *kcm)
1050{
1051 if (kcm->tx_wait_more)
1052 kcm_write_msgs(kcm);
1053}
1054
1055static ssize_t kcm_sendpage(struct socket *sock, struct page *page,
1056 int offset, size_t size, int flags)
1057
1058{
1059 struct sock *sk = sock->sk;
1060 struct kcm_sock *kcm = kcm_sk(sk);
1061 struct sk_buff *skb = NULL, *head = NULL;
1062 long timeo = sock_sndtimeo(sk, flags & MSG_DONTWAIT);
1063 bool eor;
1064 int err = 0;
1065 int i;
1066
1067 if (flags & MSG_SENDPAGE_NOTLAST)
1068 flags |= MSG_MORE;
1069
1070 /* No MSG_EOR from splice, only look at MSG_MORE */
1071 eor = !(flags & MSG_MORE);
1072
1073 lock_sock(sk);
1074
1075 sk_clear_bit(SOCKWQ_ASYNC_NOSPACE, sk);
1076
1077 err = -EPIPE;
1078 if (sk->sk_err)
1079 goto out_error;
1080
1081 if (kcm->seq_skb) {
1082 /* Previously opened message */
1083 head = kcm->seq_skb;
1084 skb = kcm_tx_msg(head)->last_skb;
1085 i = skb_shinfo(skb)->nr_frags;
1086
1087 if (skb_can_coalesce(skb, i, page, offset)) {
1088 skb_frag_size_add(&skb_shinfo(skb)->frags[i - 1], size);
1089 skb_shinfo(skb)->tx_flags |= SKBTX_SHARED_FRAG;
1090 goto coalesced;
1091 }
1092
1093 if (i >= MAX_SKB_FRAGS) {
1094 struct sk_buff *tskb;
1095
1096 tskb = alloc_skb(0, sk->sk_allocation);
1097 while (!tskb) {
1098 kcm_push(kcm);
1099 err = sk_stream_wait_memory(sk, &timeo);
1100 if (err)
1101 goto out_error;
1102 }
1103
1104 if (head == skb)
1105 skb_shinfo(head)->frag_list = tskb;
1106 else
1107 skb->next = tskb;
1108
1109 skb = tskb;
1110 skb->ip_summed = CHECKSUM_UNNECESSARY;
1111 i = 0;
1112 }
1113 } else {
1114 /* Call the sk_stream functions to manage the sndbuf mem. */
1115 if (!sk_stream_memory_free(sk)) {
1116 kcm_push(kcm);
1117 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
1118 err = sk_stream_wait_memory(sk, &timeo);
1119 if (err)
1120 goto out_error;
1121 }
1122
1123 head = alloc_skb(0, sk->sk_allocation);
1124 while (!head) {
1125 kcm_push(kcm);
1126 err = sk_stream_wait_memory(sk, &timeo);
1127 if (err)
1128 goto out_error;
1129 }
1130
1131 skb = head;
1132 i = 0;
1133 }
1134
1135 get_page(page);
1136 skb_fill_page_desc(skb, i, page, offset, size);
1137 skb_shinfo(skb)->tx_flags |= SKBTX_SHARED_FRAG;
1138
1139coalesced:
1140 skb->len += size;
1141 skb->data_len += size;
1142 skb->truesize += size;
1143 sk->sk_wmem_queued += size;
1144 sk_mem_charge(sk, size);
1145
1146 if (head != skb) {
1147 head->len += size;
1148 head->data_len += size;
1149 head->truesize += size;
1150 }
1151
1152 if (eor) {
1153 bool not_busy = skb_queue_empty(&sk->sk_write_queue);
1154
1155 /* Message complete, queue it on send buffer */
1156 __skb_queue_tail(&sk->sk_write_queue, head);
1157 kcm->seq_skb = NULL;
1158 KCM_STATS_INCR(kcm->stats.tx_msgs);
1159
1160 if (flags & MSG_BATCH) {
1161 kcm->tx_wait_more = true;
1162 } else if (kcm->tx_wait_more || not_busy) {
1163 err = kcm_write_msgs(kcm);
1164 if (err < 0) {
1165 /* We got a hard error in write_msgs but have
1166 * already queued this message. Report an error
1167 * in the socket, but don't affect return value
1168 * from sendmsg
1169 */
1170 pr_warn("KCM: Hard failure on kcm_write_msgs\n");
1171 report_csk_error(&kcm->sk, -err);
1172 }
1173 }
1174 } else {
1175 /* Message not complete, save state */
1176 kcm->seq_skb = head;
1177 kcm_tx_msg(head)->last_skb = skb;
1178 }
1179
1180 KCM_STATS_ADD(kcm->stats.tx_bytes, size);
1181
1182 release_sock(sk);
1183 return size;
1184
1185out_error:
1186 kcm_push(kcm);
1187
1188 err = sk_stream_error(sk, flags, err);
1189
1190 /* make sure we wake any epoll edge trigger waiter */
1191 if (unlikely(skb_queue_len(&sk->sk_write_queue) == 0 && err == -EAGAIN))
1192 sk->sk_write_space(sk);
1193
1194 release_sock(sk);
1195 return err;
1196}
1197
1198static int kcm_sendmsg(struct socket *sock, struct msghdr *msg, size_t len)
1199{
1200 struct sock *sk = sock->sk;
1201 struct kcm_sock *kcm = kcm_sk(sk);
1202 struct sk_buff *skb = NULL, *head = NULL;
1203 size_t copy, copied = 0;
1204 long timeo = sock_sndtimeo(sk, msg->msg_flags & MSG_DONTWAIT);
1205 int eor = (sock->type == SOCK_DGRAM) ?
1206 !(msg->msg_flags & MSG_MORE) : !!(msg->msg_flags & MSG_EOR);
1207 int err = -EPIPE;
1208
1209 lock_sock(sk);
1210
1211 /* Per tcp_sendmsg this should be in poll */
1212 sk_clear_bit(SOCKWQ_ASYNC_NOSPACE, sk);
1213
1214 if (sk->sk_err)
1215 goto out_error;
1216
1217 if (kcm->seq_skb) {
1218 /* Previously opened message */
1219 head = kcm->seq_skb;
1220 skb = kcm_tx_msg(head)->last_skb;
1221 goto start;
1222 }
1223
1224 /* Call the sk_stream functions to manage the sndbuf mem. */
1225 if (!sk_stream_memory_free(sk)) {
1226 kcm_push(kcm);
1227 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
1228 err = sk_stream_wait_memory(sk, &timeo);
1229 if (err)
1230 goto out_error;
1231 }
1232
1233 /* New message, alloc head skb */
1234 head = alloc_skb(0, sk->sk_allocation);
1235 while (!head) {
1236 kcm_push(kcm);
1237 err = sk_stream_wait_memory(sk, &timeo);
1238 if (err)
1239 goto out_error;
1240
1241 head = alloc_skb(0, sk->sk_allocation);
1242 }
1243
1244 skb = head;
1245
1246 /* Set ip_summed to CHECKSUM_UNNECESSARY to avoid calling
1247 * csum_and_copy_from_iter from skb_do_copy_data_nocache.
1248 */
1249 skb->ip_summed = CHECKSUM_UNNECESSARY;
1250
1251start:
1252 while (msg_data_left(msg)) {
1253 bool merge = true;
1254 int i = skb_shinfo(skb)->nr_frags;
1255 struct page_frag *pfrag = sk_page_frag(sk);
1256
1257 if (!sk_page_frag_refill(sk, pfrag))
1258 goto wait_for_memory;
1259
1260 if (!skb_can_coalesce(skb, i, pfrag->page,
1261 pfrag->offset)) {
1262 if (i == MAX_SKB_FRAGS) {
1263 struct sk_buff *tskb;
1264
1265 tskb = alloc_skb(0, sk->sk_allocation);
1266 if (!tskb)
1267 goto wait_for_memory;
1268
1269 if (head == skb)
1270 skb_shinfo(head)->frag_list = tskb;
1271 else
1272 skb->next = tskb;
1273
1274 skb = tskb;
1275 skb->ip_summed = CHECKSUM_UNNECESSARY;
1276 continue;
1277 }
1278 merge = false;
1279 }
1280
1281 copy = min_t(int, msg_data_left(msg),
1282 pfrag->size - pfrag->offset);
1283
1284 if (!sk_wmem_schedule(sk, copy))
1285 goto wait_for_memory;
1286
1287 err = skb_copy_to_page_nocache(sk, &msg->msg_iter, skb,
1288 pfrag->page,
1289 pfrag->offset,
1290 copy);
1291 if (err)
1292 goto out_error;
1293
1294 /* Update the skb. */
1295 if (merge) {
1296 skb_frag_size_add(&skb_shinfo(skb)->frags[i - 1], copy);
1297 } else {
1298 skb_fill_page_desc(skb, i, pfrag->page,
1299 pfrag->offset, copy);
1300 get_page(pfrag->page);
1301 }
1302
1303 pfrag->offset += copy;
1304 copied += copy;
1305 if (head != skb) {
1306 head->len += copy;
1307 head->data_len += copy;
1308 }
1309
1310 continue;
1311
1312wait_for_memory:
1313 kcm_push(kcm);
1314 err = sk_stream_wait_memory(sk, &timeo);
1315 if (err)
1316 goto out_error;
1317 }
1318
1319 if (eor) {
1320 bool not_busy = skb_queue_empty(&sk->sk_write_queue);
1321
1322 /* Message complete, queue it on send buffer */
1323 __skb_queue_tail(&sk->sk_write_queue, head);
1324 kcm->seq_skb = NULL;
1325 KCM_STATS_INCR(kcm->stats.tx_msgs);
1326
1327 if (msg->msg_flags & MSG_BATCH) {
1328 kcm->tx_wait_more = true;
1329 } else if (kcm->tx_wait_more || not_busy) {
1330 err = kcm_write_msgs(kcm);
1331 if (err < 0) {
1332 /* We got a hard error in write_msgs but have
1333 * already queued this message. Report an error
1334 * in the socket, but don't affect return value
1335 * from sendmsg
1336 */
1337 pr_warn("KCM: Hard failure on kcm_write_msgs\n");
1338 report_csk_error(&kcm->sk, -err);
1339 }
1340 }
1341 } else {
1342 /* Message not complete, save state */
1343partial_message:
1344 kcm->seq_skb = head;
1345 kcm_tx_msg(head)->last_skb = skb;
1346 }
1347
1348 KCM_STATS_ADD(kcm->stats.tx_bytes, copied);
1349
1350 release_sock(sk);
1351 return copied;
1352
1353out_error:
1354 kcm_push(kcm);
1355
1356 if (copied && sock->type == SOCK_SEQPACKET) {
1357 /* Wrote some bytes before encountering an
1358 * error, return partial success.
1359 */
1360 goto partial_message;
1361 }
1362
1363 if (head != kcm->seq_skb)
1364 kfree_skb(head);
1365
1366 err = sk_stream_error(sk, msg->msg_flags, err);
1367
1368 /* make sure we wake any epoll edge trigger waiter */
1369 if (unlikely(skb_queue_len(&sk->sk_write_queue) == 0 && err == -EAGAIN))
1370 sk->sk_write_space(sk);
1371
1372 release_sock(sk);
1373 return err;
1374}
1375
1376static struct sk_buff *kcm_wait_data(struct sock *sk, int flags,
1377 long timeo, int *err)
1378{
1379 struct sk_buff *skb;
1380
1381 while (!(skb = skb_peek(&sk->sk_receive_queue))) {
1382 if (sk->sk_err) {
1383 *err = sock_error(sk);
1384 return NULL;
1385 }
1386
1387 if (sock_flag(sk, SOCK_DONE))
1388 return NULL;
1389
1390 if ((flags & MSG_DONTWAIT) || !timeo) {
1391 *err = -EAGAIN;
1392 return NULL;
1393 }
1394
1395 sk_wait_data(sk, &timeo, NULL);
1396
1397 /* Handle signals */
1398 if (signal_pending(current)) {
1399 *err = sock_intr_errno(timeo);
1400 return NULL;
1401 }
1402 }
1403
1404 return skb;
1405}
1406
1407static int kcm_recvmsg(struct socket *sock, struct msghdr *msg,
1408 size_t len, int flags)
1409{
1410 struct sock *sk = sock->sk;
1411 struct kcm_sock *kcm = kcm_sk(sk);
1412 int err = 0;
1413 long timeo;
1414 struct kcm_rx_msg *rxm;
1415 int copied = 0;
1416 struct sk_buff *skb;
1417
1418 timeo = sock_rcvtimeo(sk, flags & MSG_DONTWAIT);
1419
1420 lock_sock(sk);
1421
1422 skb = kcm_wait_data(sk, flags, timeo, &err);
1423 if (!skb)
1424 goto out;
1425
1426 /* Okay, have a message on the receive queue */
1427
1428 rxm = kcm_rx_msg(skb);
1429
1430 if (len > rxm->full_len)
1431 len = rxm->full_len;
1432
1433 err = skb_copy_datagram_msg(skb, rxm->offset, msg, len);
1434 if (err < 0)
1435 goto out;
1436
1437 copied = len;
1438 if (likely(!(flags & MSG_PEEK))) {
1439 KCM_STATS_ADD(kcm->stats.rx_bytes, copied);
1440 if (copied < rxm->full_len) {
1441 if (sock->type == SOCK_DGRAM) {
1442 /* Truncated message */
1443 msg->msg_flags |= MSG_TRUNC;
1444 goto msg_finished;
1445 }
1446 rxm->offset += copied;
1447 rxm->full_len -= copied;
1448 } else {
1449msg_finished:
1450 /* Finished with message */
1451 msg->msg_flags |= MSG_EOR;
1452 KCM_STATS_INCR(kcm->stats.rx_msgs);
1453 skb_unlink(skb, &sk->sk_receive_queue);
1454 kfree_skb(skb);
1455 }
1456 }
1457
1458out:
1459 release_sock(sk);
1460
1461 return copied ? : err;
1462}
1463
1464static ssize_t kcm_sock_splice(struct sock *sk,
1465 struct pipe_inode_info *pipe,
1466 struct splice_pipe_desc *spd)
1467{
1468 int ret;
1469
1470 release_sock(sk);
1471 ret = splice_to_pipe(pipe, spd);
1472 lock_sock(sk);
1473
1474 return ret;
1475}
1476
1477static ssize_t kcm_splice_read(struct socket *sock, loff_t *ppos,
1478 struct pipe_inode_info *pipe, size_t len,
1479 unsigned int flags)
1480{
1481 struct sock *sk = sock->sk;
1482 struct kcm_sock *kcm = kcm_sk(sk);
1483 long timeo;
1484 struct kcm_rx_msg *rxm;
1485 int err = 0;
1486 size_t copied;
1487 struct sk_buff *skb;
1488
1489 /* Only support splice for SOCKSEQPACKET */
1490
1491 timeo = sock_rcvtimeo(sk, flags & MSG_DONTWAIT);
1492
1493 lock_sock(sk);
1494
1495 skb = kcm_wait_data(sk, flags, timeo, &err);
1496 if (!skb)
1497 goto err_out;
1498
1499 /* Okay, have a message on the receive queue */
1500
1501 rxm = kcm_rx_msg(skb);
1502
1503 if (len > rxm->full_len)
1504 len = rxm->full_len;
1505
1506 copied = skb_splice_bits(skb, sk, rxm->offset, pipe, len, flags,
1507 kcm_sock_splice);
1508 if (copied < 0) {
1509 err = copied;
1510 goto err_out;
1511 }
1512
1513 KCM_STATS_ADD(kcm->stats.rx_bytes, copied);
1514
1515 rxm->offset += copied;
1516 rxm->full_len -= copied;
1517
1518 /* We have no way to return MSG_EOR. If all the bytes have been
1519 * read we still leave the message in the receive socket buffer.
1520 * A subsequent recvmsg needs to be done to return MSG_EOR and
1521 * finish reading the message.
1522 */
1523
1524 release_sock(sk);
1525
1526 return copied;
1527
1528err_out:
1529 release_sock(sk);
1530
1531 return err;
1532}
1533
1534/* kcm sock lock held */
1535static void kcm_recv_disable(struct kcm_sock *kcm)
1536{
1537 struct kcm_mux *mux = kcm->mux;
1538
1539 if (kcm->rx_disabled)
1540 return;
1541
1542 spin_lock_bh(&mux->rx_lock);
1543
1544 kcm->rx_disabled = 1;
1545
1546 /* If a psock is reserved we'll do cleanup in unreserve */
1547 if (!kcm->rx_psock) {
1548 if (kcm->rx_wait) {
1549 list_del(&kcm->wait_rx_list);
1550 kcm->rx_wait = false;
1551 }
1552
1553 requeue_rx_msgs(mux, &kcm->sk.sk_receive_queue);
1554 }
1555
1556 spin_unlock_bh(&mux->rx_lock);
1557}
1558
1559/* kcm sock lock held */
1560static void kcm_recv_enable(struct kcm_sock *kcm)
1561{
1562 struct kcm_mux *mux = kcm->mux;
1563
1564 if (!kcm->rx_disabled)
1565 return;
1566
1567 spin_lock_bh(&mux->rx_lock);
1568
1569 kcm->rx_disabled = 0;
1570 kcm_rcv_ready(kcm);
1571
1572 spin_unlock_bh(&mux->rx_lock);
1573}
1574
1575static int kcm_setsockopt(struct socket *sock, int level, int optname,
1576 char __user *optval, unsigned int optlen)
1577{
1578 struct kcm_sock *kcm = kcm_sk(sock->sk);
1579 int val, valbool;
1580 int err = 0;
1581
1582 if (level != SOL_KCM)
1583 return -ENOPROTOOPT;
1584
1585 if (optlen < sizeof(int))
1586 return -EINVAL;
1587
1588 if (get_user(val, (int __user *)optval))
1589 return -EINVAL;
1590
1591 valbool = val ? 1 : 0;
1592
1593 switch (optname) {
1594 case KCM_RECV_DISABLE:
1595 lock_sock(&kcm->sk);
1596 if (valbool)
1597 kcm_recv_disable(kcm);
1598 else
1599 kcm_recv_enable(kcm);
1600 release_sock(&kcm->sk);
1601 break;
1602 default:
1603 err = -ENOPROTOOPT;
1604 }
1605
1606 return err;
1607}
1608
1609static int kcm_getsockopt(struct socket *sock, int level, int optname,
1610 char __user *optval, int __user *optlen)
1611{
1612 struct kcm_sock *kcm = kcm_sk(sock->sk);
1613 int val, len;
1614
1615 if (level != SOL_KCM)
1616 return -ENOPROTOOPT;
1617
1618 if (get_user(len, optlen))
1619 return -EFAULT;
1620
1621 len = min_t(unsigned int, len, sizeof(int));
1622 if (len < 0)
1623 return -EINVAL;
1624
1625 switch (optname) {
1626 case KCM_RECV_DISABLE:
1627 val = kcm->rx_disabled;
1628 break;
1629 default:
1630 return -ENOPROTOOPT;
1631 }
1632
1633 if (put_user(len, optlen))
1634 return -EFAULT;
1635 if (copy_to_user(optval, &val, len))
1636 return -EFAULT;
1637 return 0;
1638}
1639
1640static void init_kcm_sock(struct kcm_sock *kcm, struct kcm_mux *mux)
1641{
1642 struct kcm_sock *tkcm;
1643 struct list_head *head;
1644 int index = 0;
1645
1646 /* For SOCK_SEQPACKET sock type, datagram_poll checks the sk_state, so
1647 * we set sk_state, otherwise epoll_wait always returns right away with
1648 * POLLHUP
1649 */
1650 kcm->sk.sk_state = TCP_ESTABLISHED;
1651
1652 /* Add to mux's kcm sockets list */
1653 kcm->mux = mux;
1654 spin_lock_bh(&mux->lock);
1655
1656 head = &mux->kcm_socks;
1657 list_for_each_entry(tkcm, &mux->kcm_socks, kcm_sock_list) {
1658 if (tkcm->index != index)
1659 break;
1660 head = &tkcm->kcm_sock_list;
1661 index++;
1662 }
1663
1664 list_add(&kcm->kcm_sock_list, head);
1665 kcm->index = index;
1666
1667 mux->kcm_socks_cnt++;
1668 spin_unlock_bh(&mux->lock);
1669
1670 INIT_WORK(&kcm->tx_work, kcm_tx_work);
1671
1672 spin_lock_bh(&mux->rx_lock);
1673 kcm_rcv_ready(kcm);
1674 spin_unlock_bh(&mux->rx_lock);
1675}
1676
1677static void kcm_rx_msg_timeout(unsigned long arg)
1678{
1679 struct kcm_psock *psock = (struct kcm_psock *)arg;
1680
1681 /* Message assembly timed out */
1682 KCM_STATS_INCR(psock->stats.rx_msg_timeouts);
1683 kcm_abort_rx_psock(psock, ETIMEDOUT, NULL);
1684}
1685
1686static int kcm_attach(struct socket *sock, struct socket *csock,
1687 struct bpf_prog *prog)
1688{
1689 struct kcm_sock *kcm = kcm_sk(sock->sk);
1690 struct kcm_mux *mux = kcm->mux;
1691 struct sock *csk;
1692 struct kcm_psock *psock = NULL, *tpsock;
1693 struct list_head *head;
1694 int index = 0;
1695
1696 if (csock->ops->family != PF_INET &&
1697 csock->ops->family != PF_INET6)
1698 return -EINVAL;
1699
1700 csk = csock->sk;
1701 if (!csk)
1702 return -EINVAL;
1703
1704 /* Only support TCP for now */
1705 if (csk->sk_protocol != IPPROTO_TCP)
1706 return -EINVAL;
1707
1708 psock = kmem_cache_zalloc(kcm_psockp, GFP_KERNEL);
1709 if (!psock)
1710 return -ENOMEM;
1711
1712 psock->mux = mux;
1713 psock->sk = csk;
1714 psock->bpf_prog = prog;
1715
1716 setup_timer(&psock->rx_msg_timer, kcm_rx_msg_timeout,
1717 (unsigned long)psock);
1718
1719 INIT_WORK(&psock->rx_work, psock_rx_work);
1720 INIT_DELAYED_WORK(&psock->rx_delayed_work, psock_rx_delayed_work);
1721
1722 sock_hold(csk);
1723
1724 write_lock_bh(&csk->sk_callback_lock);
1725 psock->save_data_ready = csk->sk_data_ready;
1726 psock->save_write_space = csk->sk_write_space;
1727 psock->save_state_change = csk->sk_state_change;
1728 csk->sk_user_data = psock;
1729 csk->sk_data_ready = psock_tcp_data_ready;
1730 csk->sk_write_space = psock_tcp_write_space;
1731 csk->sk_state_change = psock_tcp_state_change;
1732 write_unlock_bh(&csk->sk_callback_lock);
1733
1734 /* Finished initialization, now add the psock to the MUX. */
1735 spin_lock_bh(&mux->lock);
1736 head = &mux->psocks;
1737 list_for_each_entry(tpsock, &mux->psocks, psock_list) {
1738 if (tpsock->index != index)
1739 break;
1740 head = &tpsock->psock_list;
1741 index++;
1742 }
1743
1744 list_add(&psock->psock_list, head);
1745 psock->index = index;
1746
1747 KCM_STATS_INCR(mux->stats.psock_attach);
1748 mux->psocks_cnt++;
1749 psock_now_avail(psock);
1750 spin_unlock_bh(&mux->lock);
1751
1752 /* Schedule RX work in case there are already bytes queued */
1753 queue_work(kcm_wq, &psock->rx_work);
1754
1755 return 0;
1756}
1757
1758static int kcm_attach_ioctl(struct socket *sock, struct kcm_attach *info)
1759{
1760 struct socket *csock;
1761 struct bpf_prog *prog;
1762 int err;
1763
1764 csock = sockfd_lookup(info->fd, &err);
1765 if (!csock)
1766 return -ENOENT;
1767
1768 prog = bpf_prog_get(info->bpf_fd);
1769 if (IS_ERR(prog)) {
1770 err = PTR_ERR(prog);
1771 goto out;
1772 }
1773
1774 if (prog->type != BPF_PROG_TYPE_SOCKET_FILTER) {
1775 bpf_prog_put(prog);
1776 err = -EINVAL;
1777 goto out;
1778 }
1779
1780 err = kcm_attach(sock, csock, prog);
1781 if (err) {
1782 bpf_prog_put(prog);
1783 goto out;
1784 }
1785
1786 /* Keep reference on file also */
1787
1788 return 0;
1789out:
1790 fput(csock->file);
1791 return err;
1792}
1793
1794static void kcm_unattach(struct kcm_psock *psock)
1795{
1796 struct sock *csk = psock->sk;
1797 struct kcm_mux *mux = psock->mux;
1798
1799 /* Stop getting callbacks from TCP socket. After this there should
1800 * be no way to reserve a kcm for this psock.
1801 */
1802 write_lock_bh(&csk->sk_callback_lock);
1803 csk->sk_user_data = NULL;
1804 csk->sk_data_ready = psock->save_data_ready;
1805 csk->sk_write_space = psock->save_write_space;
1806 csk->sk_state_change = psock->save_state_change;
1807 psock->rx_stopped = 1;
1808
1809 if (WARN_ON(psock->rx_kcm)) {
1810 write_unlock_bh(&csk->sk_callback_lock);
1811 return;
1812 }
1813
1814 spin_lock_bh(&mux->rx_lock);
1815
1816 /* Stop receiver activities. After this point psock should not be
1817 * able to get onto ready list either through callbacks or work.
1818 */
1819 if (psock->ready_rx_msg) {
1820 list_del(&psock->psock_ready_list);
1821 kfree_skb(psock->ready_rx_msg);
1822 psock->ready_rx_msg = NULL;
1823 KCM_STATS_INCR(mux->stats.rx_ready_drops);
1824 }
1825
1826 spin_unlock_bh(&mux->rx_lock);
1827
1828 write_unlock_bh(&csk->sk_callback_lock);
1829
1830 del_timer_sync(&psock->rx_msg_timer);
1831 cancel_work_sync(&psock->rx_work);
1832 cancel_delayed_work_sync(&psock->rx_delayed_work);
1833
1834 bpf_prog_put(psock->bpf_prog);
1835
1836 kfree_skb(psock->rx_skb_head);
1837 psock->rx_skb_head = NULL;
1838
1839 spin_lock_bh(&mux->lock);
1840
1841 aggregate_psock_stats(&psock->stats, &mux->aggregate_psock_stats);
1842
1843 KCM_STATS_INCR(mux->stats.psock_unattach);
1844
1845 if (psock->tx_kcm) {
1846 /* psock was reserved. Just mark it finished and we will clean
1847 * up in the kcm paths, we need kcm lock which can not be
1848 * acquired here.
1849 */
1850 KCM_STATS_INCR(mux->stats.psock_unattach_rsvd);
1851 spin_unlock_bh(&mux->lock);
1852
1853 /* We are unattaching a socket that is reserved. Abort the
1854 * socket since we may be out of sync in sending on it. We need
1855 * to do this without the mux lock.
1856 */
1857 kcm_abort_tx_psock(psock, EPIPE, false);
1858
1859 spin_lock_bh(&mux->lock);
1860 if (!psock->tx_kcm) {
1861 /* psock now unreserved in window mux was unlocked */
1862 goto no_reserved;
1863 }
1864 psock->done = 1;
1865
1866 /* Commit done before queuing work to process it */
1867 smp_mb();
1868
1869 /* Queue tx work to make sure psock->done is handled */
1870 queue_work(kcm_wq, &psock->tx_kcm->tx_work);
1871 spin_unlock_bh(&mux->lock);
1872 } else {
1873no_reserved:
1874 if (!psock->tx_stopped)
1875 list_del(&psock->psock_avail_list);
1876 list_del(&psock->psock_list);
1877 mux->psocks_cnt--;
1878 spin_unlock_bh(&mux->lock);
1879
1880 sock_put(csk);
1881 fput(csk->sk_socket->file);
1882 kmem_cache_free(kcm_psockp, psock);
1883 }
1884}
1885
1886static int kcm_unattach_ioctl(struct socket *sock, struct kcm_unattach *info)
1887{
1888 struct kcm_sock *kcm = kcm_sk(sock->sk);
1889 struct kcm_mux *mux = kcm->mux;
1890 struct kcm_psock *psock;
1891 struct socket *csock;
1892 struct sock *csk;
1893 int err;
1894
1895 csock = sockfd_lookup(info->fd, &err);
1896 if (!csock)
1897 return -ENOENT;
1898
1899 csk = csock->sk;
1900 if (!csk) {
1901 err = -EINVAL;
1902 goto out;
1903 }
1904
1905 err = -ENOENT;
1906
1907 spin_lock_bh(&mux->lock);
1908
1909 list_for_each_entry(psock, &mux->psocks, psock_list) {
1910 if (psock->sk != csk)
1911 continue;
1912
1913 /* Found the matching psock */
1914
1915 if (psock->unattaching || WARN_ON(psock->done)) {
1916 err = -EALREADY;
1917 break;
1918 }
1919
1920 psock->unattaching = 1;
1921
1922 spin_unlock_bh(&mux->lock);
1923
1924 kcm_unattach(psock);
1925
1926 err = 0;
1927 goto out;
1928 }
1929
1930 spin_unlock_bh(&mux->lock);
1931
1932out:
1933 fput(csock->file);
1934 return err;
1935}
1936
1937static struct proto kcm_proto = {
1938 .name = "KCM",
1939 .owner = THIS_MODULE,
1940 .obj_size = sizeof(struct kcm_sock),
1941};
1942
1943/* Clone a kcm socket. */
1944static int kcm_clone(struct socket *osock, struct kcm_clone *info,
1945 struct socket **newsockp)
1946{
1947 struct socket *newsock;
1948 struct sock *newsk;
1949 struct file *newfile;
1950 int err, newfd;
1951
1952 err = -ENFILE;
1953 newsock = sock_alloc();
1954 if (!newsock)
1955 goto out;
1956
1957 newsock->type = osock->type;
1958 newsock->ops = osock->ops;
1959
1960 __module_get(newsock->ops->owner);
1961
1962 newfd = get_unused_fd_flags(0);
1963 if (unlikely(newfd < 0)) {
1964 err = newfd;
1965 goto out_fd_fail;
1966 }
1967
1968 newfile = sock_alloc_file(newsock, 0, osock->sk->sk_prot_creator->name);
1969 if (unlikely(IS_ERR(newfile))) {
1970 err = PTR_ERR(newfile);
1971 goto out_sock_alloc_fail;
1972 }
1973
1974 newsk = sk_alloc(sock_net(osock->sk), PF_KCM, GFP_KERNEL,
1975 &kcm_proto, true);
1976 if (!newsk) {
1977 err = -ENOMEM;
1978 goto out_sk_alloc_fail;
1979 }
1980
1981 sock_init_data(newsock, newsk);
1982 init_kcm_sock(kcm_sk(newsk), kcm_sk(osock->sk)->mux);
1983
1984 fd_install(newfd, newfile);
1985 *newsockp = newsock;
1986 info->fd = newfd;
1987
1988 return 0;
1989
1990out_sk_alloc_fail:
1991 fput(newfile);
1992out_sock_alloc_fail:
1993 put_unused_fd(newfd);
1994out_fd_fail:
1995 sock_release(newsock);
1996out:
1997 return err;
1998}
1999
2000static int kcm_ioctl(struct socket *sock, unsigned int cmd, unsigned long arg)
2001{
2002 int err;
2003
2004 switch (cmd) {
2005 case SIOCKCMATTACH: {
2006 struct kcm_attach info;
2007
2008 if (copy_from_user(&info, (void __user *)arg, sizeof(info)))
2009 err = -EFAULT;
2010
2011 err = kcm_attach_ioctl(sock, &info);
2012
2013 break;
2014 }
2015 case SIOCKCMUNATTACH: {
2016 struct kcm_unattach info;
2017
2018 if (copy_from_user(&info, (void __user *)arg, sizeof(info)))
2019 err = -EFAULT;
2020
2021 err = kcm_unattach_ioctl(sock, &info);
2022
2023 break;
2024 }
2025 case SIOCKCMCLONE: {
2026 struct kcm_clone info;
2027 struct socket *newsock = NULL;
2028
2029 if (copy_from_user(&info, (void __user *)arg, sizeof(info)))
2030 err = -EFAULT;
2031
2032 err = kcm_clone(sock, &info, &newsock);
2033
2034 if (!err) {
2035 if (copy_to_user((void __user *)arg, &info,
2036 sizeof(info))) {
2037 err = -EFAULT;
2038 sock_release(newsock);
2039 }
2040 }
2041
2042 break;
2043 }
2044 default:
2045 err = -ENOIOCTLCMD;
2046 break;
2047 }
2048
2049 return err;
2050}
2051
2052static void free_mux(struct rcu_head *rcu)
2053{
2054 struct kcm_mux *mux = container_of(rcu,
2055 struct kcm_mux, rcu);
2056
2057 kmem_cache_free(kcm_muxp, mux);
2058}
2059
2060static void release_mux(struct kcm_mux *mux)
2061{
2062 struct kcm_net *knet = mux->knet;
2063 struct kcm_psock *psock, *tmp_psock;
2064
2065 /* Release psocks */
2066 list_for_each_entry_safe(psock, tmp_psock,
2067 &mux->psocks, psock_list) {
2068 if (!WARN_ON(psock->unattaching))
2069 kcm_unattach(psock);
2070 }
2071
2072 if (WARN_ON(mux->psocks_cnt))
2073 return;
2074
2075 __skb_queue_purge(&mux->rx_hold_queue);
2076
2077 mutex_lock(&knet->mutex);
2078 aggregate_mux_stats(&mux->stats, &knet->aggregate_mux_stats);
2079 aggregate_psock_stats(&mux->aggregate_psock_stats,
2080 &knet->aggregate_psock_stats);
2081 list_del_rcu(&mux->kcm_mux_list);
2082 knet->count--;
2083 mutex_unlock(&knet->mutex);
2084
2085 call_rcu(&mux->rcu, free_mux);
2086}
2087
2088static void kcm_done(struct kcm_sock *kcm)
2089{
2090 struct kcm_mux *mux = kcm->mux;
2091 struct sock *sk = &kcm->sk;
2092 int socks_cnt;
2093
2094 spin_lock_bh(&mux->rx_lock);
2095 if (kcm->rx_psock) {
2096 /* Cleanup in unreserve_rx_kcm */
2097 WARN_ON(kcm->done);
2098 kcm->rx_disabled = 1;
2099 kcm->done = 1;
2100 spin_unlock_bh(&mux->rx_lock);
2101 return;
2102 }
2103
2104 if (kcm->rx_wait) {
2105 list_del(&kcm->wait_rx_list);
2106 kcm->rx_wait = false;
2107 }
2108 /* Move any pending receive messages to other kcm sockets */
2109 requeue_rx_msgs(mux, &sk->sk_receive_queue);
2110
2111 spin_unlock_bh(&mux->rx_lock);
2112
2113 if (WARN_ON(sk_rmem_alloc_get(sk)))
2114 return;
2115
2116 /* Detach from MUX */
2117 spin_lock_bh(&mux->lock);
2118
2119 list_del(&kcm->kcm_sock_list);
2120 mux->kcm_socks_cnt--;
2121 socks_cnt = mux->kcm_socks_cnt;
2122
2123 spin_unlock_bh(&mux->lock);
2124
2125 if (!socks_cnt) {
2126 /* We are done with the mux now. */
2127 release_mux(mux);
2128 }
2129
2130 WARN_ON(kcm->rx_wait);
2131
2132 sock_put(&kcm->sk);
2133}
2134
2135/* Called by kcm_release to close a KCM socket.
2136 * If this is the last KCM socket on the MUX, destroy the MUX.
2137 */
2138static int kcm_release(struct socket *sock)
2139{
2140 struct sock *sk = sock->sk;
2141 struct kcm_sock *kcm;
2142 struct kcm_mux *mux;
2143 struct kcm_psock *psock;
2144
2145 if (!sk)
2146 return 0;
2147
2148 kcm = kcm_sk(sk);
2149 mux = kcm->mux;
2150
2151 sock_orphan(sk);
2152 kfree_skb(kcm->seq_skb);
2153
2154 lock_sock(sk);
2155 /* Purge queue under lock to avoid race condition with tx_work trying
2156 * to act when queue is nonempty. If tx_work runs after this point
2157 * it will just return.
2158 */
2159 __skb_queue_purge(&sk->sk_write_queue);
2160 release_sock(sk);
2161
2162 spin_lock_bh(&mux->lock);
2163 if (kcm->tx_wait) {
2164 /* Take of tx_wait list, after this point there should be no way
2165 * that a psock will be assigned to this kcm.
2166 */
2167 list_del(&kcm->wait_psock_list);
2168 kcm->tx_wait = false;
2169 }
2170 spin_unlock_bh(&mux->lock);
2171
2172 /* Cancel work. After this point there should be no outside references
2173 * to the kcm socket.
2174 */
2175 cancel_work_sync(&kcm->tx_work);
2176
2177 lock_sock(sk);
2178 psock = kcm->tx_psock;
2179 if (psock) {
2180 /* A psock was reserved, so we need to kill it since it
2181 * may already have some bytes queued from a message. We
2182 * need to do this after removing kcm from tx_wait list.
2183 */
2184 kcm_abort_tx_psock(psock, EPIPE, false);
2185 unreserve_psock(kcm);
2186 }
2187 release_sock(sk);
2188
2189 WARN_ON(kcm->tx_wait);
2190 WARN_ON(kcm->tx_psock);
2191
2192 sock->sk = NULL;
2193
2194 kcm_done(kcm);
2195
2196 return 0;
2197}
2198
2199static const struct proto_ops kcm_dgram_ops = {
2200 .family = PF_KCM,
2201 .owner = THIS_MODULE,
2202 .release = kcm_release,
2203 .bind = sock_no_bind,
2204 .connect = sock_no_connect,
2205 .socketpair = sock_no_socketpair,
2206 .accept = sock_no_accept,
2207 .getname = sock_no_getname,
2208 .poll = datagram_poll,
2209 .ioctl = kcm_ioctl,
2210 .listen = sock_no_listen,
2211 .shutdown = sock_no_shutdown,
2212 .setsockopt = kcm_setsockopt,
2213 .getsockopt = kcm_getsockopt,
2214 .sendmsg = kcm_sendmsg,
2215 .recvmsg = kcm_recvmsg,
2216 .mmap = sock_no_mmap,
2217 .sendpage = kcm_sendpage,
2218};
2219
2220static const struct proto_ops kcm_seqpacket_ops = {
2221 .family = PF_KCM,
2222 .owner = THIS_MODULE,
2223 .release = kcm_release,
2224 .bind = sock_no_bind,
2225 .connect = sock_no_connect,
2226 .socketpair = sock_no_socketpair,
2227 .accept = sock_no_accept,
2228 .getname = sock_no_getname,
2229 .poll = datagram_poll,
2230 .ioctl = kcm_ioctl,
2231 .listen = sock_no_listen,
2232 .shutdown = sock_no_shutdown,
2233 .setsockopt = kcm_setsockopt,
2234 .getsockopt = kcm_getsockopt,
2235 .sendmsg = kcm_sendmsg,
2236 .recvmsg = kcm_recvmsg,
2237 .mmap = sock_no_mmap,
2238 .sendpage = kcm_sendpage,
2239 .splice_read = kcm_splice_read,
2240};
2241
2242/* Create proto operation for kcm sockets */
2243static int kcm_create(struct net *net, struct socket *sock,
2244 int protocol, int kern)
2245{
2246 struct kcm_net *knet = net_generic(net, kcm_net_id);
2247 struct sock *sk;
2248 struct kcm_mux *mux;
2249
2250 switch (sock->type) {
2251 case SOCK_DGRAM:
2252 sock->ops = &kcm_dgram_ops;
2253 break;
2254 case SOCK_SEQPACKET:
2255 sock->ops = &kcm_seqpacket_ops;
2256 break;
2257 default:
2258 return -ESOCKTNOSUPPORT;
2259 }
2260
2261 if (protocol != KCMPROTO_CONNECTED)
2262 return -EPROTONOSUPPORT;
2263
2264 sk = sk_alloc(net, PF_KCM, GFP_KERNEL, &kcm_proto, kern);
2265 if (!sk)
2266 return -ENOMEM;
2267
2268 /* Allocate a kcm mux, shared between KCM sockets */
2269 mux = kmem_cache_zalloc(kcm_muxp, GFP_KERNEL);
2270 if (!mux) {
2271 sk_free(sk);
2272 return -ENOMEM;
2273 }
2274
2275 spin_lock_init(&mux->lock);
2276 spin_lock_init(&mux->rx_lock);
2277 INIT_LIST_HEAD(&mux->kcm_socks);
2278 INIT_LIST_HEAD(&mux->kcm_rx_waiters);
2279 INIT_LIST_HEAD(&mux->kcm_tx_waiters);
2280
2281 INIT_LIST_HEAD(&mux->psocks);
2282 INIT_LIST_HEAD(&mux->psocks_ready);
2283 INIT_LIST_HEAD(&mux->psocks_avail);
2284
2285 mux->knet = knet;
2286
2287 /* Add new MUX to list */
2288 mutex_lock(&knet->mutex);
2289 list_add_rcu(&mux->kcm_mux_list, &knet->mux_list);
2290 knet->count++;
2291 mutex_unlock(&knet->mutex);
2292
2293 skb_queue_head_init(&mux->rx_hold_queue);
2294
2295 /* Init KCM socket */
2296 sock_init_data(sock, sk);
2297 init_kcm_sock(kcm_sk(sk), mux);
2298
2299 return 0;
2300}
2301
2302static struct net_proto_family kcm_family_ops = {
2303 .family = PF_KCM,
2304 .create = kcm_create,
2305 .owner = THIS_MODULE,
2306};
2307
2308static __net_init int kcm_init_net(struct net *net)
2309{
2310 struct kcm_net *knet = net_generic(net, kcm_net_id);
2311
2312 INIT_LIST_HEAD_RCU(&knet->mux_list);
2313 mutex_init(&knet->mutex);
2314
2315 return 0;
2316}
2317
2318static __net_exit void kcm_exit_net(struct net *net)
2319{
2320 struct kcm_net *knet = net_generic(net, kcm_net_id);
2321
2322 /* All KCM sockets should be closed at this point, which should mean
2323 * that all multiplexors and psocks have been destroyed.
2324 */
2325 WARN_ON(!list_empty(&knet->mux_list));
2326}
2327
2328static struct pernet_operations kcm_net_ops = {
2329 .init = kcm_init_net,
2330 .exit = kcm_exit_net,
2331 .id = &kcm_net_id,
2332 .size = sizeof(struct kcm_net),
2333};
2334
2335static int __init kcm_init(void)
2336{
2337 int err = -ENOMEM;
2338
2339 kcm_muxp = kmem_cache_create("kcm_mux_cache",
2340 sizeof(struct kcm_mux), 0,
2341 SLAB_HWCACHE_ALIGN | SLAB_PANIC, NULL);
2342 if (!kcm_muxp)
2343 goto fail;
2344
2345 kcm_psockp = kmem_cache_create("kcm_psock_cache",
2346 sizeof(struct kcm_psock), 0,
2347 SLAB_HWCACHE_ALIGN | SLAB_PANIC, NULL);
2348 if (!kcm_psockp)
2349 goto fail;
2350
2351 kcm_wq = create_singlethread_workqueue("kkcmd");
2352 if (!kcm_wq)
2353 goto fail;
2354
2355 err = proto_register(&kcm_proto, 1);
2356 if (err)
2357 goto fail;
2358
2359 err = sock_register(&kcm_family_ops);
2360 if (err)
2361 goto sock_register_fail;
2362
2363 err = register_pernet_device(&kcm_net_ops);
2364 if (err)
2365 goto net_ops_fail;
2366
2367 err = kcm_proc_init();
2368 if (err)
2369 goto proc_init_fail;
2370
2371 return 0;
2372
2373proc_init_fail:
2374 unregister_pernet_device(&kcm_net_ops);
2375
2376net_ops_fail:
2377 sock_unregister(PF_KCM);
2378
2379sock_register_fail:
2380 proto_unregister(&kcm_proto);
2381
2382fail:
2383 kmem_cache_destroy(kcm_muxp);
2384 kmem_cache_destroy(kcm_psockp);
2385
2386 if (kcm_wq)
2387 destroy_workqueue(kcm_wq);
2388
2389 return err;
2390}
2391
2392static void __exit kcm_exit(void)
2393{
2394 kcm_proc_exit();
2395 unregister_pernet_device(&kcm_net_ops);
2396 sock_unregister(PF_KCM);
2397 proto_unregister(&kcm_proto);
2398 destroy_workqueue(kcm_wq);
2399
2400 kmem_cache_destroy(kcm_muxp);
2401 kmem_cache_destroy(kcm_psockp);
2402}
2403
2404module_init(kcm_init);
2405module_exit(kcm_exit);
2406
2407MODULE_LICENSE("GPL");
2408MODULE_ALIAS_NETPROTO(PF_KCM);
2409
1// SPDX-License-Identifier: GPL-2.0-only
2/*
3 * Kernel Connection Multiplexor
4 *
5 * Copyright (c) 2016 Tom Herbert <tom@herbertland.com>
6 */
7
8#include <linux/bpf.h>
9#include <linux/errno.h>
10#include <linux/errqueue.h>
11#include <linux/file.h>
12#include <linux/filter.h>
13#include <linux/in.h>
14#include <linux/kernel.h>
15#include <linux/module.h>
16#include <linux/net.h>
17#include <linux/netdevice.h>
18#include <linux/poll.h>
19#include <linux/rculist.h>
20#include <linux/skbuff.h>
21#include <linux/socket.h>
22#include <linux/uaccess.h>
23#include <linux/workqueue.h>
24#include <linux/syscalls.h>
25#include <linux/sched/signal.h>
26
27#include <net/kcm.h>
28#include <net/netns/generic.h>
29#include <net/sock.h>
30#include <uapi/linux/kcm.h>
31#include <trace/events/sock.h>
32
33unsigned int kcm_net_id;
34
35static struct kmem_cache *kcm_psockp __read_mostly;
36static struct kmem_cache *kcm_muxp __read_mostly;
37static struct workqueue_struct *kcm_wq;
38
39static inline struct kcm_sock *kcm_sk(const struct sock *sk)
40{
41 return (struct kcm_sock *)sk;
42}
43
44static inline struct kcm_tx_msg *kcm_tx_msg(struct sk_buff *skb)
45{
46 return (struct kcm_tx_msg *)skb->cb;
47}
48
49static void report_csk_error(struct sock *csk, int err)
50{
51 csk->sk_err = EPIPE;
52 sk_error_report(csk);
53}
54
55static void kcm_abort_tx_psock(struct kcm_psock *psock, int err,
56 bool wakeup_kcm)
57{
58 struct sock *csk = psock->sk;
59 struct kcm_mux *mux = psock->mux;
60
61 /* Unrecoverable error in transmit */
62
63 spin_lock_bh(&mux->lock);
64
65 if (psock->tx_stopped) {
66 spin_unlock_bh(&mux->lock);
67 return;
68 }
69
70 psock->tx_stopped = 1;
71 KCM_STATS_INCR(psock->stats.tx_aborts);
72
73 if (!psock->tx_kcm) {
74 /* Take off psocks_avail list */
75 list_del(&psock->psock_avail_list);
76 } else if (wakeup_kcm) {
77 /* In this case psock is being aborted while outside of
78 * write_msgs and psock is reserved. Schedule tx_work
79 * to handle the failure there. Need to commit tx_stopped
80 * before queuing work.
81 */
82 smp_mb();
83
84 queue_work(kcm_wq, &psock->tx_kcm->tx_work);
85 }
86
87 spin_unlock_bh(&mux->lock);
88
89 /* Report error on lower socket */
90 report_csk_error(csk, err);
91}
92
93/* RX mux lock held. */
94static void kcm_update_rx_mux_stats(struct kcm_mux *mux,
95 struct kcm_psock *psock)
96{
97 STRP_STATS_ADD(mux->stats.rx_bytes,
98 psock->strp.stats.bytes -
99 psock->saved_rx_bytes);
100 mux->stats.rx_msgs +=
101 psock->strp.stats.msgs - psock->saved_rx_msgs;
102 psock->saved_rx_msgs = psock->strp.stats.msgs;
103 psock->saved_rx_bytes = psock->strp.stats.bytes;
104}
105
106static void kcm_update_tx_mux_stats(struct kcm_mux *mux,
107 struct kcm_psock *psock)
108{
109 KCM_STATS_ADD(mux->stats.tx_bytes,
110 psock->stats.tx_bytes - psock->saved_tx_bytes);
111 mux->stats.tx_msgs +=
112 psock->stats.tx_msgs - psock->saved_tx_msgs;
113 psock->saved_tx_msgs = psock->stats.tx_msgs;
114 psock->saved_tx_bytes = psock->stats.tx_bytes;
115}
116
117static int kcm_queue_rcv_skb(struct sock *sk, struct sk_buff *skb);
118
119/* KCM is ready to receive messages on its queue-- either the KCM is new or
120 * has become unblocked after being blocked on full socket buffer. Queue any
121 * pending ready messages on a psock. RX mux lock held.
122 */
123static void kcm_rcv_ready(struct kcm_sock *kcm)
124{
125 struct kcm_mux *mux = kcm->mux;
126 struct kcm_psock *psock;
127 struct sk_buff *skb;
128
129 if (unlikely(kcm->rx_wait || kcm->rx_psock || kcm->rx_disabled))
130 return;
131
132 while (unlikely((skb = __skb_dequeue(&mux->rx_hold_queue)))) {
133 if (kcm_queue_rcv_skb(&kcm->sk, skb)) {
134 /* Assuming buffer limit has been reached */
135 skb_queue_head(&mux->rx_hold_queue, skb);
136 WARN_ON(!sk_rmem_alloc_get(&kcm->sk));
137 return;
138 }
139 }
140
141 while (!list_empty(&mux->psocks_ready)) {
142 psock = list_first_entry(&mux->psocks_ready, struct kcm_psock,
143 psock_ready_list);
144
145 if (kcm_queue_rcv_skb(&kcm->sk, psock->ready_rx_msg)) {
146 /* Assuming buffer limit has been reached */
147 WARN_ON(!sk_rmem_alloc_get(&kcm->sk));
148 return;
149 }
150
151 /* Consumed the ready message on the psock. Schedule rx_work to
152 * get more messages.
153 */
154 list_del(&psock->psock_ready_list);
155 psock->ready_rx_msg = NULL;
156 /* Commit clearing of ready_rx_msg for queuing work */
157 smp_mb();
158
159 strp_unpause(&psock->strp);
160 strp_check_rcv(&psock->strp);
161 }
162
163 /* Buffer limit is okay now, add to ready list */
164 list_add_tail(&kcm->wait_rx_list,
165 &kcm->mux->kcm_rx_waiters);
166 /* paired with lockless reads in kcm_rfree() */
167 WRITE_ONCE(kcm->rx_wait, true);
168}
169
170static void kcm_rfree(struct sk_buff *skb)
171{
172 struct sock *sk = skb->sk;
173 struct kcm_sock *kcm = kcm_sk(sk);
174 struct kcm_mux *mux = kcm->mux;
175 unsigned int len = skb->truesize;
176
177 sk_mem_uncharge(sk, len);
178 atomic_sub(len, &sk->sk_rmem_alloc);
179
180 /* For reading rx_wait and rx_psock without holding lock */
181 smp_mb__after_atomic();
182
183 if (!READ_ONCE(kcm->rx_wait) && !READ_ONCE(kcm->rx_psock) &&
184 sk_rmem_alloc_get(sk) < sk->sk_rcvlowat) {
185 spin_lock_bh(&mux->rx_lock);
186 kcm_rcv_ready(kcm);
187 spin_unlock_bh(&mux->rx_lock);
188 }
189}
190
191static int kcm_queue_rcv_skb(struct sock *sk, struct sk_buff *skb)
192{
193 struct sk_buff_head *list = &sk->sk_receive_queue;
194
195 if (atomic_read(&sk->sk_rmem_alloc) >= sk->sk_rcvbuf)
196 return -ENOMEM;
197
198 if (!sk_rmem_schedule(sk, skb, skb->truesize))
199 return -ENOBUFS;
200
201 skb->dev = NULL;
202
203 skb_orphan(skb);
204 skb->sk = sk;
205 skb->destructor = kcm_rfree;
206 atomic_add(skb->truesize, &sk->sk_rmem_alloc);
207 sk_mem_charge(sk, skb->truesize);
208
209 skb_queue_tail(list, skb);
210
211 if (!sock_flag(sk, SOCK_DEAD))
212 sk->sk_data_ready(sk);
213
214 return 0;
215}
216
217/* Requeue received messages for a kcm socket to other kcm sockets. This is
218 * called with a kcm socket is receive disabled.
219 * RX mux lock held.
220 */
221static void requeue_rx_msgs(struct kcm_mux *mux, struct sk_buff_head *head)
222{
223 struct sk_buff *skb;
224 struct kcm_sock *kcm;
225
226 while ((skb = skb_dequeue(head))) {
227 /* Reset destructor to avoid calling kcm_rcv_ready */
228 skb->destructor = sock_rfree;
229 skb_orphan(skb);
230try_again:
231 if (list_empty(&mux->kcm_rx_waiters)) {
232 skb_queue_tail(&mux->rx_hold_queue, skb);
233 continue;
234 }
235
236 kcm = list_first_entry(&mux->kcm_rx_waiters,
237 struct kcm_sock, wait_rx_list);
238
239 if (kcm_queue_rcv_skb(&kcm->sk, skb)) {
240 /* Should mean socket buffer full */
241 list_del(&kcm->wait_rx_list);
242 /* paired with lockless reads in kcm_rfree() */
243 WRITE_ONCE(kcm->rx_wait, false);
244
245 /* Commit rx_wait to read in kcm_free */
246 smp_wmb();
247
248 goto try_again;
249 }
250 }
251}
252
253/* Lower sock lock held */
254static struct kcm_sock *reserve_rx_kcm(struct kcm_psock *psock,
255 struct sk_buff *head)
256{
257 struct kcm_mux *mux = psock->mux;
258 struct kcm_sock *kcm;
259
260 WARN_ON(psock->ready_rx_msg);
261
262 if (psock->rx_kcm)
263 return psock->rx_kcm;
264
265 spin_lock_bh(&mux->rx_lock);
266
267 if (psock->rx_kcm) {
268 spin_unlock_bh(&mux->rx_lock);
269 return psock->rx_kcm;
270 }
271
272 kcm_update_rx_mux_stats(mux, psock);
273
274 if (list_empty(&mux->kcm_rx_waiters)) {
275 psock->ready_rx_msg = head;
276 strp_pause(&psock->strp);
277 list_add_tail(&psock->psock_ready_list,
278 &mux->psocks_ready);
279 spin_unlock_bh(&mux->rx_lock);
280 return NULL;
281 }
282
283 kcm = list_first_entry(&mux->kcm_rx_waiters,
284 struct kcm_sock, wait_rx_list);
285 list_del(&kcm->wait_rx_list);
286 /* paired with lockless reads in kcm_rfree() */
287 WRITE_ONCE(kcm->rx_wait, false);
288
289 psock->rx_kcm = kcm;
290 /* paired with lockless reads in kcm_rfree() */
291 WRITE_ONCE(kcm->rx_psock, psock);
292
293 spin_unlock_bh(&mux->rx_lock);
294
295 return kcm;
296}
297
298static void kcm_done(struct kcm_sock *kcm);
299
300static void kcm_done_work(struct work_struct *w)
301{
302 kcm_done(container_of(w, struct kcm_sock, done_work));
303}
304
305/* Lower sock held */
306static void unreserve_rx_kcm(struct kcm_psock *psock,
307 bool rcv_ready)
308{
309 struct kcm_sock *kcm = psock->rx_kcm;
310 struct kcm_mux *mux = psock->mux;
311
312 if (!kcm)
313 return;
314
315 spin_lock_bh(&mux->rx_lock);
316
317 psock->rx_kcm = NULL;
318 /* paired with lockless reads in kcm_rfree() */
319 WRITE_ONCE(kcm->rx_psock, NULL);
320
321 /* Commit kcm->rx_psock before sk_rmem_alloc_get to sync with
322 * kcm_rfree
323 */
324 smp_mb();
325
326 if (unlikely(kcm->done)) {
327 spin_unlock_bh(&mux->rx_lock);
328
329 /* Need to run kcm_done in a task since we need to qcquire
330 * callback locks which may already be held here.
331 */
332 INIT_WORK(&kcm->done_work, kcm_done_work);
333 schedule_work(&kcm->done_work);
334 return;
335 }
336
337 if (unlikely(kcm->rx_disabled)) {
338 requeue_rx_msgs(mux, &kcm->sk.sk_receive_queue);
339 } else if (rcv_ready || unlikely(!sk_rmem_alloc_get(&kcm->sk))) {
340 /* Check for degenerative race with rx_wait that all
341 * data was dequeued (accounted for in kcm_rfree).
342 */
343 kcm_rcv_ready(kcm);
344 }
345 spin_unlock_bh(&mux->rx_lock);
346}
347
348/* Lower sock lock held */
349static void psock_data_ready(struct sock *sk)
350{
351 struct kcm_psock *psock;
352
353 trace_sk_data_ready(sk);
354
355 read_lock_bh(&sk->sk_callback_lock);
356
357 psock = (struct kcm_psock *)sk->sk_user_data;
358 if (likely(psock))
359 strp_data_ready(&psock->strp);
360
361 read_unlock_bh(&sk->sk_callback_lock);
362}
363
364/* Called with lower sock held */
365static void kcm_rcv_strparser(struct strparser *strp, struct sk_buff *skb)
366{
367 struct kcm_psock *psock = container_of(strp, struct kcm_psock, strp);
368 struct kcm_sock *kcm;
369
370try_queue:
371 kcm = reserve_rx_kcm(psock, skb);
372 if (!kcm) {
373 /* Unable to reserve a KCM, message is held in psock and strp
374 * is paused.
375 */
376 return;
377 }
378
379 if (kcm_queue_rcv_skb(&kcm->sk, skb)) {
380 /* Should mean socket buffer full */
381 unreserve_rx_kcm(psock, false);
382 goto try_queue;
383 }
384}
385
386static int kcm_parse_func_strparser(struct strparser *strp, struct sk_buff *skb)
387{
388 struct kcm_psock *psock = container_of(strp, struct kcm_psock, strp);
389 struct bpf_prog *prog = psock->bpf_prog;
390 int res;
391
392 res = bpf_prog_run_pin_on_cpu(prog, skb);
393 return res;
394}
395
396static int kcm_read_sock_done(struct strparser *strp, int err)
397{
398 struct kcm_psock *psock = container_of(strp, struct kcm_psock, strp);
399
400 unreserve_rx_kcm(psock, true);
401
402 return err;
403}
404
405static void psock_state_change(struct sock *sk)
406{
407 /* TCP only does a EPOLLIN for a half close. Do a EPOLLHUP here
408 * since application will normally not poll with EPOLLIN
409 * on the TCP sockets.
410 */
411
412 report_csk_error(sk, EPIPE);
413}
414
415static void psock_write_space(struct sock *sk)
416{
417 struct kcm_psock *psock;
418 struct kcm_mux *mux;
419 struct kcm_sock *kcm;
420
421 read_lock_bh(&sk->sk_callback_lock);
422
423 psock = (struct kcm_psock *)sk->sk_user_data;
424 if (unlikely(!psock))
425 goto out;
426 mux = psock->mux;
427
428 spin_lock_bh(&mux->lock);
429
430 /* Check if the socket is reserved so someone is waiting for sending. */
431 kcm = psock->tx_kcm;
432 if (kcm && !unlikely(kcm->tx_stopped))
433 queue_work(kcm_wq, &kcm->tx_work);
434
435 spin_unlock_bh(&mux->lock);
436out:
437 read_unlock_bh(&sk->sk_callback_lock);
438}
439
440static void unreserve_psock(struct kcm_sock *kcm);
441
442/* kcm sock is locked. */
443static struct kcm_psock *reserve_psock(struct kcm_sock *kcm)
444{
445 struct kcm_mux *mux = kcm->mux;
446 struct kcm_psock *psock;
447
448 psock = kcm->tx_psock;
449
450 smp_rmb(); /* Must read tx_psock before tx_wait */
451
452 if (psock) {
453 WARN_ON(kcm->tx_wait);
454 if (unlikely(psock->tx_stopped))
455 unreserve_psock(kcm);
456 else
457 return kcm->tx_psock;
458 }
459
460 spin_lock_bh(&mux->lock);
461
462 /* Check again under lock to see if psock was reserved for this
463 * psock via psock_unreserve.
464 */
465 psock = kcm->tx_psock;
466 if (unlikely(psock)) {
467 WARN_ON(kcm->tx_wait);
468 spin_unlock_bh(&mux->lock);
469 return kcm->tx_psock;
470 }
471
472 if (!list_empty(&mux->psocks_avail)) {
473 psock = list_first_entry(&mux->psocks_avail,
474 struct kcm_psock,
475 psock_avail_list);
476 list_del(&psock->psock_avail_list);
477 if (kcm->tx_wait) {
478 list_del(&kcm->wait_psock_list);
479 kcm->tx_wait = false;
480 }
481 kcm->tx_psock = psock;
482 psock->tx_kcm = kcm;
483 KCM_STATS_INCR(psock->stats.reserved);
484 } else if (!kcm->tx_wait) {
485 list_add_tail(&kcm->wait_psock_list,
486 &mux->kcm_tx_waiters);
487 kcm->tx_wait = true;
488 }
489
490 spin_unlock_bh(&mux->lock);
491
492 return psock;
493}
494
495/* mux lock held */
496static void psock_now_avail(struct kcm_psock *psock)
497{
498 struct kcm_mux *mux = psock->mux;
499 struct kcm_sock *kcm;
500
501 if (list_empty(&mux->kcm_tx_waiters)) {
502 list_add_tail(&psock->psock_avail_list,
503 &mux->psocks_avail);
504 } else {
505 kcm = list_first_entry(&mux->kcm_tx_waiters,
506 struct kcm_sock,
507 wait_psock_list);
508 list_del(&kcm->wait_psock_list);
509 kcm->tx_wait = false;
510 psock->tx_kcm = kcm;
511
512 /* Commit before changing tx_psock since that is read in
513 * reserve_psock before queuing work.
514 */
515 smp_mb();
516
517 kcm->tx_psock = psock;
518 KCM_STATS_INCR(psock->stats.reserved);
519 queue_work(kcm_wq, &kcm->tx_work);
520 }
521}
522
523/* kcm sock is locked. */
524static void unreserve_psock(struct kcm_sock *kcm)
525{
526 struct kcm_psock *psock;
527 struct kcm_mux *mux = kcm->mux;
528
529 spin_lock_bh(&mux->lock);
530
531 psock = kcm->tx_psock;
532
533 if (WARN_ON(!psock)) {
534 spin_unlock_bh(&mux->lock);
535 return;
536 }
537
538 smp_rmb(); /* Read tx_psock before tx_wait */
539
540 kcm_update_tx_mux_stats(mux, psock);
541
542 WARN_ON(kcm->tx_wait);
543
544 kcm->tx_psock = NULL;
545 psock->tx_kcm = NULL;
546 KCM_STATS_INCR(psock->stats.unreserved);
547
548 if (unlikely(psock->tx_stopped)) {
549 if (psock->done) {
550 /* Deferred free */
551 list_del(&psock->psock_list);
552 mux->psocks_cnt--;
553 sock_put(psock->sk);
554 fput(psock->sk->sk_socket->file);
555 kmem_cache_free(kcm_psockp, psock);
556 }
557
558 /* Don't put back on available list */
559
560 spin_unlock_bh(&mux->lock);
561
562 return;
563 }
564
565 psock_now_avail(psock);
566
567 spin_unlock_bh(&mux->lock);
568}
569
570static void kcm_report_tx_retry(struct kcm_sock *kcm)
571{
572 struct kcm_mux *mux = kcm->mux;
573
574 spin_lock_bh(&mux->lock);
575 KCM_STATS_INCR(mux->stats.tx_retries);
576 spin_unlock_bh(&mux->lock);
577}
578
579/* Write any messages ready on the kcm socket. Called with kcm sock lock
580 * held. Return bytes actually sent or error.
581 */
582static int kcm_write_msgs(struct kcm_sock *kcm)
583{
584 unsigned int total_sent = 0;
585 struct sock *sk = &kcm->sk;
586 struct kcm_psock *psock;
587 struct sk_buff *head;
588 int ret = 0;
589
590 kcm->tx_wait_more = false;
591 psock = kcm->tx_psock;
592 if (unlikely(psock && psock->tx_stopped)) {
593 /* A reserved psock was aborted asynchronously. Unreserve
594 * it and we'll retry the message.
595 */
596 unreserve_psock(kcm);
597 kcm_report_tx_retry(kcm);
598 if (skb_queue_empty(&sk->sk_write_queue))
599 return 0;
600
601 kcm_tx_msg(skb_peek(&sk->sk_write_queue))->started_tx = false;
602 }
603
604retry:
605 while ((head = skb_peek(&sk->sk_write_queue))) {
606 struct msghdr msg = {
607 .msg_flags = MSG_DONTWAIT | MSG_SPLICE_PAGES,
608 };
609 struct kcm_tx_msg *txm = kcm_tx_msg(head);
610 struct sk_buff *skb;
611 unsigned int msize;
612 int i;
613
614 if (!txm->started_tx) {
615 psock = reserve_psock(kcm);
616 if (!psock)
617 goto out;
618 skb = head;
619 txm->frag_offset = 0;
620 txm->sent = 0;
621 txm->started_tx = true;
622 } else {
623 if (WARN_ON(!psock)) {
624 ret = -EINVAL;
625 goto out;
626 }
627 skb = txm->frag_skb;
628 }
629
630 if (WARN_ON(!skb_shinfo(skb)->nr_frags) ||
631 WARN_ON_ONCE(!skb_frag_page(&skb_shinfo(skb)->frags[0]))) {
632 ret = -EINVAL;
633 goto out;
634 }
635
636 msize = 0;
637 for (i = 0; i < skb_shinfo(skb)->nr_frags; i++)
638 msize += skb_frag_size(&skb_shinfo(skb)->frags[i]);
639
640 iov_iter_bvec(&msg.msg_iter, ITER_SOURCE,
641 (const struct bio_vec *)skb_shinfo(skb)->frags,
642 skb_shinfo(skb)->nr_frags, msize);
643 iov_iter_advance(&msg.msg_iter, txm->frag_offset);
644
645 do {
646 ret = sock_sendmsg(psock->sk->sk_socket, &msg);
647 if (ret <= 0) {
648 if (ret == -EAGAIN) {
649 /* Save state to try again when there's
650 * write space on the socket
651 */
652 txm->frag_skb = skb;
653 ret = 0;
654 goto out;
655 }
656
657 /* Hard failure in sending message, abort this
658 * psock since it has lost framing
659 * synchronization and retry sending the
660 * message from the beginning.
661 */
662 kcm_abort_tx_psock(psock, ret ? -ret : EPIPE,
663 true);
664 unreserve_psock(kcm);
665 psock = NULL;
666
667 txm->started_tx = false;
668 kcm_report_tx_retry(kcm);
669 ret = 0;
670 goto retry;
671 }
672
673 txm->sent += ret;
674 txm->frag_offset += ret;
675 KCM_STATS_ADD(psock->stats.tx_bytes, ret);
676 } while (msg.msg_iter.count > 0);
677
678 if (skb == head) {
679 if (skb_has_frag_list(skb)) {
680 txm->frag_skb = skb_shinfo(skb)->frag_list;
681 txm->frag_offset = 0;
682 continue;
683 }
684 } else if (skb->next) {
685 txm->frag_skb = skb->next;
686 txm->frag_offset = 0;
687 continue;
688 }
689
690 /* Successfully sent the whole packet, account for it. */
691 sk->sk_wmem_queued -= txm->sent;
692 total_sent += txm->sent;
693 skb_dequeue(&sk->sk_write_queue);
694 kfree_skb(head);
695 KCM_STATS_INCR(psock->stats.tx_msgs);
696 }
697out:
698 if (!head) {
699 /* Done with all queued messages. */
700 WARN_ON(!skb_queue_empty(&sk->sk_write_queue));
701 if (psock)
702 unreserve_psock(kcm);
703 }
704
705 /* Check if write space is available */
706 sk->sk_write_space(sk);
707
708 return total_sent ? : ret;
709}
710
711static void kcm_tx_work(struct work_struct *w)
712{
713 struct kcm_sock *kcm = container_of(w, struct kcm_sock, tx_work);
714 struct sock *sk = &kcm->sk;
715 int err;
716
717 lock_sock(sk);
718
719 /* Primarily for SOCK_DGRAM sockets, also handle asynchronous tx
720 * aborts
721 */
722 err = kcm_write_msgs(kcm);
723 if (err < 0) {
724 /* Hard failure in write, report error on KCM socket */
725 pr_warn("KCM: Hard failure on kcm_write_msgs %d\n", err);
726 report_csk_error(&kcm->sk, -err);
727 goto out;
728 }
729
730 /* Primarily for SOCK_SEQPACKET sockets */
731 if (likely(sk->sk_socket) &&
732 test_bit(SOCK_NOSPACE, &sk->sk_socket->flags)) {
733 clear_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
734 sk->sk_write_space(sk);
735 }
736
737out:
738 release_sock(sk);
739}
740
741static void kcm_push(struct kcm_sock *kcm)
742{
743 if (kcm->tx_wait_more)
744 kcm_write_msgs(kcm);
745}
746
747static int kcm_sendmsg(struct socket *sock, struct msghdr *msg, size_t len)
748{
749 struct sock *sk = sock->sk;
750 struct kcm_sock *kcm = kcm_sk(sk);
751 struct sk_buff *skb = NULL, *head = NULL;
752 size_t copy, copied = 0;
753 long timeo = sock_sndtimeo(sk, msg->msg_flags & MSG_DONTWAIT);
754 int eor = (sock->type == SOCK_DGRAM) ?
755 !(msg->msg_flags & MSG_MORE) : !!(msg->msg_flags & MSG_EOR);
756 int err = -EPIPE;
757
758 mutex_lock(&kcm->tx_mutex);
759 lock_sock(sk);
760
761 /* Per tcp_sendmsg this should be in poll */
762 sk_clear_bit(SOCKWQ_ASYNC_NOSPACE, sk);
763
764 if (sk->sk_err)
765 goto out_error;
766
767 if (kcm->seq_skb) {
768 /* Previously opened message */
769 head = kcm->seq_skb;
770 skb = kcm_tx_msg(head)->last_skb;
771 goto start;
772 }
773
774 /* Call the sk_stream functions to manage the sndbuf mem. */
775 if (!sk_stream_memory_free(sk)) {
776 kcm_push(kcm);
777 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
778 err = sk_stream_wait_memory(sk, &timeo);
779 if (err)
780 goto out_error;
781 }
782
783 if (msg_data_left(msg)) {
784 /* New message, alloc head skb */
785 head = alloc_skb(0, sk->sk_allocation);
786 while (!head) {
787 kcm_push(kcm);
788 err = sk_stream_wait_memory(sk, &timeo);
789 if (err)
790 goto out_error;
791
792 head = alloc_skb(0, sk->sk_allocation);
793 }
794
795 skb = head;
796
797 /* Set ip_summed to CHECKSUM_UNNECESSARY to avoid calling
798 * csum_and_copy_from_iter from skb_do_copy_data_nocache.
799 */
800 skb->ip_summed = CHECKSUM_UNNECESSARY;
801 }
802
803start:
804 while (msg_data_left(msg)) {
805 bool merge = true;
806 int i = skb_shinfo(skb)->nr_frags;
807 struct page_frag *pfrag = sk_page_frag(sk);
808
809 if (!sk_page_frag_refill(sk, pfrag))
810 goto wait_for_memory;
811
812 if (!skb_can_coalesce(skb, i, pfrag->page,
813 pfrag->offset)) {
814 if (i == MAX_SKB_FRAGS) {
815 struct sk_buff *tskb;
816
817 tskb = alloc_skb(0, sk->sk_allocation);
818 if (!tskb)
819 goto wait_for_memory;
820
821 if (head == skb)
822 skb_shinfo(head)->frag_list = tskb;
823 else
824 skb->next = tskb;
825
826 skb = tskb;
827 skb->ip_summed = CHECKSUM_UNNECESSARY;
828 continue;
829 }
830 merge = false;
831 }
832
833 if (msg->msg_flags & MSG_SPLICE_PAGES) {
834 copy = msg_data_left(msg);
835 if (!sk_wmem_schedule(sk, copy))
836 goto wait_for_memory;
837
838 err = skb_splice_from_iter(skb, &msg->msg_iter, copy,
839 sk->sk_allocation);
840 if (err < 0) {
841 if (err == -EMSGSIZE)
842 goto wait_for_memory;
843 goto out_error;
844 }
845
846 copy = err;
847 skb_shinfo(skb)->flags |= SKBFL_SHARED_FRAG;
848 sk_wmem_queued_add(sk, copy);
849 sk_mem_charge(sk, copy);
850
851 if (head != skb)
852 head->truesize += copy;
853 } else {
854 copy = min_t(int, msg_data_left(msg),
855 pfrag->size - pfrag->offset);
856 if (!sk_wmem_schedule(sk, copy))
857 goto wait_for_memory;
858
859 err = skb_copy_to_page_nocache(sk, &msg->msg_iter, skb,
860 pfrag->page,
861 pfrag->offset,
862 copy);
863 if (err)
864 goto out_error;
865
866 /* Update the skb. */
867 if (merge) {
868 skb_frag_size_add(
869 &skb_shinfo(skb)->frags[i - 1], copy);
870 } else {
871 skb_fill_page_desc(skb, i, pfrag->page,
872 pfrag->offset, copy);
873 get_page(pfrag->page);
874 }
875
876 pfrag->offset += copy;
877 }
878
879 copied += copy;
880 if (head != skb) {
881 head->len += copy;
882 head->data_len += copy;
883 }
884
885 continue;
886
887wait_for_memory:
888 kcm_push(kcm);
889 err = sk_stream_wait_memory(sk, &timeo);
890 if (err)
891 goto out_error;
892 }
893
894 if (eor) {
895 bool not_busy = skb_queue_empty(&sk->sk_write_queue);
896
897 if (head) {
898 /* Message complete, queue it on send buffer */
899 __skb_queue_tail(&sk->sk_write_queue, head);
900 kcm->seq_skb = NULL;
901 KCM_STATS_INCR(kcm->stats.tx_msgs);
902 }
903
904 if (msg->msg_flags & MSG_BATCH) {
905 kcm->tx_wait_more = true;
906 } else if (kcm->tx_wait_more || not_busy) {
907 err = kcm_write_msgs(kcm);
908 if (err < 0) {
909 /* We got a hard error in write_msgs but have
910 * already queued this message. Report an error
911 * in the socket, but don't affect return value
912 * from sendmsg
913 */
914 pr_warn("KCM: Hard failure on kcm_write_msgs\n");
915 report_csk_error(&kcm->sk, -err);
916 }
917 }
918 } else {
919 /* Message not complete, save state */
920partial_message:
921 if (head) {
922 kcm->seq_skb = head;
923 kcm_tx_msg(head)->last_skb = skb;
924 }
925 }
926
927 KCM_STATS_ADD(kcm->stats.tx_bytes, copied);
928
929 release_sock(sk);
930 mutex_unlock(&kcm->tx_mutex);
931 return copied;
932
933out_error:
934 kcm_push(kcm);
935
936 if (sock->type == SOCK_SEQPACKET) {
937 /* Wrote some bytes before encountering an
938 * error, return partial success.
939 */
940 if (copied)
941 goto partial_message;
942 if (head != kcm->seq_skb)
943 kfree_skb(head);
944 } else {
945 kfree_skb(head);
946 kcm->seq_skb = NULL;
947 }
948
949 err = sk_stream_error(sk, msg->msg_flags, err);
950
951 /* make sure we wake any epoll edge trigger waiter */
952 if (unlikely(skb_queue_len(&sk->sk_write_queue) == 0 && err == -EAGAIN))
953 sk->sk_write_space(sk);
954
955 release_sock(sk);
956 mutex_unlock(&kcm->tx_mutex);
957 return err;
958}
959
960static void kcm_splice_eof(struct socket *sock)
961{
962 struct sock *sk = sock->sk;
963 struct kcm_sock *kcm = kcm_sk(sk);
964
965 if (skb_queue_empty_lockless(&sk->sk_write_queue))
966 return;
967
968 lock_sock(sk);
969 kcm_write_msgs(kcm);
970 release_sock(sk);
971}
972
973static int kcm_recvmsg(struct socket *sock, struct msghdr *msg,
974 size_t len, int flags)
975{
976 struct sock *sk = sock->sk;
977 struct kcm_sock *kcm = kcm_sk(sk);
978 int err = 0;
979 struct strp_msg *stm;
980 int copied = 0;
981 struct sk_buff *skb;
982
983 skb = skb_recv_datagram(sk, flags, &err);
984 if (!skb)
985 goto out;
986
987 /* Okay, have a message on the receive queue */
988
989 stm = strp_msg(skb);
990
991 if (len > stm->full_len)
992 len = stm->full_len;
993
994 err = skb_copy_datagram_msg(skb, stm->offset, msg, len);
995 if (err < 0)
996 goto out;
997
998 copied = len;
999 if (likely(!(flags & MSG_PEEK))) {
1000 KCM_STATS_ADD(kcm->stats.rx_bytes, copied);
1001 if (copied < stm->full_len) {
1002 if (sock->type == SOCK_DGRAM) {
1003 /* Truncated message */
1004 msg->msg_flags |= MSG_TRUNC;
1005 goto msg_finished;
1006 }
1007 stm->offset += copied;
1008 stm->full_len -= copied;
1009 } else {
1010msg_finished:
1011 /* Finished with message */
1012 msg->msg_flags |= MSG_EOR;
1013 KCM_STATS_INCR(kcm->stats.rx_msgs);
1014 }
1015 }
1016
1017out:
1018 skb_free_datagram(sk, skb);
1019 return copied ? : err;
1020}
1021
1022static ssize_t kcm_splice_read(struct socket *sock, loff_t *ppos,
1023 struct pipe_inode_info *pipe, size_t len,
1024 unsigned int flags)
1025{
1026 struct sock *sk = sock->sk;
1027 struct kcm_sock *kcm = kcm_sk(sk);
1028 struct strp_msg *stm;
1029 int err = 0;
1030 ssize_t copied;
1031 struct sk_buff *skb;
1032
1033 /* Only support splice for SOCKSEQPACKET */
1034
1035 skb = skb_recv_datagram(sk, flags, &err);
1036 if (!skb)
1037 goto err_out;
1038
1039 /* Okay, have a message on the receive queue */
1040
1041 stm = strp_msg(skb);
1042
1043 if (len > stm->full_len)
1044 len = stm->full_len;
1045
1046 copied = skb_splice_bits(skb, sk, stm->offset, pipe, len, flags);
1047 if (copied < 0) {
1048 err = copied;
1049 goto err_out;
1050 }
1051
1052 KCM_STATS_ADD(kcm->stats.rx_bytes, copied);
1053
1054 stm->offset += copied;
1055 stm->full_len -= copied;
1056
1057 /* We have no way to return MSG_EOR. If all the bytes have been
1058 * read we still leave the message in the receive socket buffer.
1059 * A subsequent recvmsg needs to be done to return MSG_EOR and
1060 * finish reading the message.
1061 */
1062
1063 skb_free_datagram(sk, skb);
1064 return copied;
1065
1066err_out:
1067 skb_free_datagram(sk, skb);
1068 return err;
1069}
1070
1071/* kcm sock lock held */
1072static void kcm_recv_disable(struct kcm_sock *kcm)
1073{
1074 struct kcm_mux *mux = kcm->mux;
1075
1076 if (kcm->rx_disabled)
1077 return;
1078
1079 spin_lock_bh(&mux->rx_lock);
1080
1081 kcm->rx_disabled = 1;
1082
1083 /* If a psock is reserved we'll do cleanup in unreserve */
1084 if (!kcm->rx_psock) {
1085 if (kcm->rx_wait) {
1086 list_del(&kcm->wait_rx_list);
1087 /* paired with lockless reads in kcm_rfree() */
1088 WRITE_ONCE(kcm->rx_wait, false);
1089 }
1090
1091 requeue_rx_msgs(mux, &kcm->sk.sk_receive_queue);
1092 }
1093
1094 spin_unlock_bh(&mux->rx_lock);
1095}
1096
1097/* kcm sock lock held */
1098static void kcm_recv_enable(struct kcm_sock *kcm)
1099{
1100 struct kcm_mux *mux = kcm->mux;
1101
1102 if (!kcm->rx_disabled)
1103 return;
1104
1105 spin_lock_bh(&mux->rx_lock);
1106
1107 kcm->rx_disabled = 0;
1108 kcm_rcv_ready(kcm);
1109
1110 spin_unlock_bh(&mux->rx_lock);
1111}
1112
1113static int kcm_setsockopt(struct socket *sock, int level, int optname,
1114 sockptr_t optval, unsigned int optlen)
1115{
1116 struct kcm_sock *kcm = kcm_sk(sock->sk);
1117 int val, valbool;
1118 int err = 0;
1119
1120 if (level != SOL_KCM)
1121 return -ENOPROTOOPT;
1122
1123 if (optlen < sizeof(int))
1124 return -EINVAL;
1125
1126 if (copy_from_sockptr(&val, optval, sizeof(int)))
1127 return -EFAULT;
1128
1129 valbool = val ? 1 : 0;
1130
1131 switch (optname) {
1132 case KCM_RECV_DISABLE:
1133 lock_sock(&kcm->sk);
1134 if (valbool)
1135 kcm_recv_disable(kcm);
1136 else
1137 kcm_recv_enable(kcm);
1138 release_sock(&kcm->sk);
1139 break;
1140 default:
1141 err = -ENOPROTOOPT;
1142 }
1143
1144 return err;
1145}
1146
1147static int kcm_getsockopt(struct socket *sock, int level, int optname,
1148 char __user *optval, int __user *optlen)
1149{
1150 struct kcm_sock *kcm = kcm_sk(sock->sk);
1151 int val, len;
1152
1153 if (level != SOL_KCM)
1154 return -ENOPROTOOPT;
1155
1156 if (get_user(len, optlen))
1157 return -EFAULT;
1158
1159 if (len < 0)
1160 return -EINVAL;
1161
1162 len = min_t(unsigned int, len, sizeof(int));
1163
1164 switch (optname) {
1165 case KCM_RECV_DISABLE:
1166 val = kcm->rx_disabled;
1167 break;
1168 default:
1169 return -ENOPROTOOPT;
1170 }
1171
1172 if (put_user(len, optlen))
1173 return -EFAULT;
1174 if (copy_to_user(optval, &val, len))
1175 return -EFAULT;
1176 return 0;
1177}
1178
1179static void init_kcm_sock(struct kcm_sock *kcm, struct kcm_mux *mux)
1180{
1181 struct kcm_sock *tkcm;
1182 struct list_head *head;
1183 int index = 0;
1184
1185 /* For SOCK_SEQPACKET sock type, datagram_poll checks the sk_state, so
1186 * we set sk_state, otherwise epoll_wait always returns right away with
1187 * EPOLLHUP
1188 */
1189 kcm->sk.sk_state = TCP_ESTABLISHED;
1190
1191 /* Add to mux's kcm sockets list */
1192 kcm->mux = mux;
1193 spin_lock_bh(&mux->lock);
1194
1195 head = &mux->kcm_socks;
1196 list_for_each_entry(tkcm, &mux->kcm_socks, kcm_sock_list) {
1197 if (tkcm->index != index)
1198 break;
1199 head = &tkcm->kcm_sock_list;
1200 index++;
1201 }
1202
1203 list_add(&kcm->kcm_sock_list, head);
1204 kcm->index = index;
1205
1206 mux->kcm_socks_cnt++;
1207 spin_unlock_bh(&mux->lock);
1208
1209 INIT_WORK(&kcm->tx_work, kcm_tx_work);
1210 mutex_init(&kcm->tx_mutex);
1211
1212 spin_lock_bh(&mux->rx_lock);
1213 kcm_rcv_ready(kcm);
1214 spin_unlock_bh(&mux->rx_lock);
1215}
1216
1217static int kcm_attach(struct socket *sock, struct socket *csock,
1218 struct bpf_prog *prog)
1219{
1220 struct kcm_sock *kcm = kcm_sk(sock->sk);
1221 struct kcm_mux *mux = kcm->mux;
1222 struct sock *csk;
1223 struct kcm_psock *psock = NULL, *tpsock;
1224 struct list_head *head;
1225 int index = 0;
1226 static const struct strp_callbacks cb = {
1227 .rcv_msg = kcm_rcv_strparser,
1228 .parse_msg = kcm_parse_func_strparser,
1229 .read_sock_done = kcm_read_sock_done,
1230 };
1231 int err = 0;
1232
1233 csk = csock->sk;
1234 if (!csk)
1235 return -EINVAL;
1236
1237 lock_sock(csk);
1238
1239 /* Only allow TCP sockets to be attached for now */
1240 if ((csk->sk_family != AF_INET && csk->sk_family != AF_INET6) ||
1241 csk->sk_protocol != IPPROTO_TCP) {
1242 err = -EOPNOTSUPP;
1243 goto out;
1244 }
1245
1246 /* Don't allow listeners or closed sockets */
1247 if (csk->sk_state == TCP_LISTEN || csk->sk_state == TCP_CLOSE) {
1248 err = -EOPNOTSUPP;
1249 goto out;
1250 }
1251
1252 psock = kmem_cache_zalloc(kcm_psockp, GFP_KERNEL);
1253 if (!psock) {
1254 err = -ENOMEM;
1255 goto out;
1256 }
1257
1258 psock->mux = mux;
1259 psock->sk = csk;
1260 psock->bpf_prog = prog;
1261
1262 write_lock_bh(&csk->sk_callback_lock);
1263
1264 /* Check if sk_user_data is already by KCM or someone else.
1265 * Must be done under lock to prevent race conditions.
1266 */
1267 if (csk->sk_user_data) {
1268 write_unlock_bh(&csk->sk_callback_lock);
1269 kmem_cache_free(kcm_psockp, psock);
1270 err = -EALREADY;
1271 goto out;
1272 }
1273
1274 err = strp_init(&psock->strp, csk, &cb);
1275 if (err) {
1276 write_unlock_bh(&csk->sk_callback_lock);
1277 kmem_cache_free(kcm_psockp, psock);
1278 goto out;
1279 }
1280
1281 psock->save_data_ready = csk->sk_data_ready;
1282 psock->save_write_space = csk->sk_write_space;
1283 psock->save_state_change = csk->sk_state_change;
1284 csk->sk_user_data = psock;
1285 csk->sk_data_ready = psock_data_ready;
1286 csk->sk_write_space = psock_write_space;
1287 csk->sk_state_change = psock_state_change;
1288
1289 write_unlock_bh(&csk->sk_callback_lock);
1290
1291 sock_hold(csk);
1292
1293 /* Finished initialization, now add the psock to the MUX. */
1294 spin_lock_bh(&mux->lock);
1295 head = &mux->psocks;
1296 list_for_each_entry(tpsock, &mux->psocks, psock_list) {
1297 if (tpsock->index != index)
1298 break;
1299 head = &tpsock->psock_list;
1300 index++;
1301 }
1302
1303 list_add(&psock->psock_list, head);
1304 psock->index = index;
1305
1306 KCM_STATS_INCR(mux->stats.psock_attach);
1307 mux->psocks_cnt++;
1308 psock_now_avail(psock);
1309 spin_unlock_bh(&mux->lock);
1310
1311 /* Schedule RX work in case there are already bytes queued */
1312 strp_check_rcv(&psock->strp);
1313
1314out:
1315 release_sock(csk);
1316
1317 return err;
1318}
1319
1320static int kcm_attach_ioctl(struct socket *sock, struct kcm_attach *info)
1321{
1322 struct socket *csock;
1323 struct bpf_prog *prog;
1324 int err;
1325
1326 csock = sockfd_lookup(info->fd, &err);
1327 if (!csock)
1328 return -ENOENT;
1329
1330 prog = bpf_prog_get_type(info->bpf_fd, BPF_PROG_TYPE_SOCKET_FILTER);
1331 if (IS_ERR(prog)) {
1332 err = PTR_ERR(prog);
1333 goto out;
1334 }
1335
1336 err = kcm_attach(sock, csock, prog);
1337 if (err) {
1338 bpf_prog_put(prog);
1339 goto out;
1340 }
1341
1342 /* Keep reference on file also */
1343
1344 return 0;
1345out:
1346 sockfd_put(csock);
1347 return err;
1348}
1349
1350static void kcm_unattach(struct kcm_psock *psock)
1351{
1352 struct sock *csk = psock->sk;
1353 struct kcm_mux *mux = psock->mux;
1354
1355 lock_sock(csk);
1356
1357 /* Stop getting callbacks from TCP socket. After this there should
1358 * be no way to reserve a kcm for this psock.
1359 */
1360 write_lock_bh(&csk->sk_callback_lock);
1361 csk->sk_user_data = NULL;
1362 csk->sk_data_ready = psock->save_data_ready;
1363 csk->sk_write_space = psock->save_write_space;
1364 csk->sk_state_change = psock->save_state_change;
1365 strp_stop(&psock->strp);
1366
1367 if (WARN_ON(psock->rx_kcm)) {
1368 write_unlock_bh(&csk->sk_callback_lock);
1369 release_sock(csk);
1370 return;
1371 }
1372
1373 spin_lock_bh(&mux->rx_lock);
1374
1375 /* Stop receiver activities. After this point psock should not be
1376 * able to get onto ready list either through callbacks or work.
1377 */
1378 if (psock->ready_rx_msg) {
1379 list_del(&psock->psock_ready_list);
1380 kfree_skb(psock->ready_rx_msg);
1381 psock->ready_rx_msg = NULL;
1382 KCM_STATS_INCR(mux->stats.rx_ready_drops);
1383 }
1384
1385 spin_unlock_bh(&mux->rx_lock);
1386
1387 write_unlock_bh(&csk->sk_callback_lock);
1388
1389 /* Call strp_done without sock lock */
1390 release_sock(csk);
1391 strp_done(&psock->strp);
1392 lock_sock(csk);
1393
1394 bpf_prog_put(psock->bpf_prog);
1395
1396 spin_lock_bh(&mux->lock);
1397
1398 aggregate_psock_stats(&psock->stats, &mux->aggregate_psock_stats);
1399 save_strp_stats(&psock->strp, &mux->aggregate_strp_stats);
1400
1401 KCM_STATS_INCR(mux->stats.psock_unattach);
1402
1403 if (psock->tx_kcm) {
1404 /* psock was reserved. Just mark it finished and we will clean
1405 * up in the kcm paths, we need kcm lock which can not be
1406 * acquired here.
1407 */
1408 KCM_STATS_INCR(mux->stats.psock_unattach_rsvd);
1409 spin_unlock_bh(&mux->lock);
1410
1411 /* We are unattaching a socket that is reserved. Abort the
1412 * socket since we may be out of sync in sending on it. We need
1413 * to do this without the mux lock.
1414 */
1415 kcm_abort_tx_psock(psock, EPIPE, false);
1416
1417 spin_lock_bh(&mux->lock);
1418 if (!psock->tx_kcm) {
1419 /* psock now unreserved in window mux was unlocked */
1420 goto no_reserved;
1421 }
1422 psock->done = 1;
1423
1424 /* Commit done before queuing work to process it */
1425 smp_mb();
1426
1427 /* Queue tx work to make sure psock->done is handled */
1428 queue_work(kcm_wq, &psock->tx_kcm->tx_work);
1429 spin_unlock_bh(&mux->lock);
1430 } else {
1431no_reserved:
1432 if (!psock->tx_stopped)
1433 list_del(&psock->psock_avail_list);
1434 list_del(&psock->psock_list);
1435 mux->psocks_cnt--;
1436 spin_unlock_bh(&mux->lock);
1437
1438 sock_put(csk);
1439 fput(csk->sk_socket->file);
1440 kmem_cache_free(kcm_psockp, psock);
1441 }
1442
1443 release_sock(csk);
1444}
1445
1446static int kcm_unattach_ioctl(struct socket *sock, struct kcm_unattach *info)
1447{
1448 struct kcm_sock *kcm = kcm_sk(sock->sk);
1449 struct kcm_mux *mux = kcm->mux;
1450 struct kcm_psock *psock;
1451 struct socket *csock;
1452 struct sock *csk;
1453 int err;
1454
1455 csock = sockfd_lookup(info->fd, &err);
1456 if (!csock)
1457 return -ENOENT;
1458
1459 csk = csock->sk;
1460 if (!csk) {
1461 err = -EINVAL;
1462 goto out;
1463 }
1464
1465 err = -ENOENT;
1466
1467 spin_lock_bh(&mux->lock);
1468
1469 list_for_each_entry(psock, &mux->psocks, psock_list) {
1470 if (psock->sk != csk)
1471 continue;
1472
1473 /* Found the matching psock */
1474
1475 if (psock->unattaching || WARN_ON(psock->done)) {
1476 err = -EALREADY;
1477 break;
1478 }
1479
1480 psock->unattaching = 1;
1481
1482 spin_unlock_bh(&mux->lock);
1483
1484 /* Lower socket lock should already be held */
1485 kcm_unattach(psock);
1486
1487 err = 0;
1488 goto out;
1489 }
1490
1491 spin_unlock_bh(&mux->lock);
1492
1493out:
1494 sockfd_put(csock);
1495 return err;
1496}
1497
1498static struct proto kcm_proto = {
1499 .name = "KCM",
1500 .owner = THIS_MODULE,
1501 .obj_size = sizeof(struct kcm_sock),
1502};
1503
1504/* Clone a kcm socket. */
1505static struct file *kcm_clone(struct socket *osock)
1506{
1507 struct socket *newsock;
1508 struct sock *newsk;
1509
1510 newsock = sock_alloc();
1511 if (!newsock)
1512 return ERR_PTR(-ENFILE);
1513
1514 newsock->type = osock->type;
1515 newsock->ops = osock->ops;
1516
1517 __module_get(newsock->ops->owner);
1518
1519 newsk = sk_alloc(sock_net(osock->sk), PF_KCM, GFP_KERNEL,
1520 &kcm_proto, false);
1521 if (!newsk) {
1522 sock_release(newsock);
1523 return ERR_PTR(-ENOMEM);
1524 }
1525 sock_init_data(newsock, newsk);
1526 init_kcm_sock(kcm_sk(newsk), kcm_sk(osock->sk)->mux);
1527
1528 return sock_alloc_file(newsock, 0, osock->sk->sk_prot_creator->name);
1529}
1530
1531static int kcm_ioctl(struct socket *sock, unsigned int cmd, unsigned long arg)
1532{
1533 int err;
1534
1535 switch (cmd) {
1536 case SIOCKCMATTACH: {
1537 struct kcm_attach info;
1538
1539 if (copy_from_user(&info, (void __user *)arg, sizeof(info)))
1540 return -EFAULT;
1541
1542 err = kcm_attach_ioctl(sock, &info);
1543
1544 break;
1545 }
1546 case SIOCKCMUNATTACH: {
1547 struct kcm_unattach info;
1548
1549 if (copy_from_user(&info, (void __user *)arg, sizeof(info)))
1550 return -EFAULT;
1551
1552 err = kcm_unattach_ioctl(sock, &info);
1553
1554 break;
1555 }
1556 case SIOCKCMCLONE: {
1557 struct kcm_clone info;
1558 struct file *file;
1559
1560 info.fd = get_unused_fd_flags(0);
1561 if (unlikely(info.fd < 0))
1562 return info.fd;
1563
1564 file = kcm_clone(sock);
1565 if (IS_ERR(file)) {
1566 put_unused_fd(info.fd);
1567 return PTR_ERR(file);
1568 }
1569 if (copy_to_user((void __user *)arg, &info,
1570 sizeof(info))) {
1571 put_unused_fd(info.fd);
1572 fput(file);
1573 return -EFAULT;
1574 }
1575 fd_install(info.fd, file);
1576 err = 0;
1577 break;
1578 }
1579 default:
1580 err = -ENOIOCTLCMD;
1581 break;
1582 }
1583
1584 return err;
1585}
1586
1587static void release_mux(struct kcm_mux *mux)
1588{
1589 struct kcm_net *knet = mux->knet;
1590 struct kcm_psock *psock, *tmp_psock;
1591
1592 /* Release psocks */
1593 list_for_each_entry_safe(psock, tmp_psock,
1594 &mux->psocks, psock_list) {
1595 if (!WARN_ON(psock->unattaching))
1596 kcm_unattach(psock);
1597 }
1598
1599 if (WARN_ON(mux->psocks_cnt))
1600 return;
1601
1602 __skb_queue_purge(&mux->rx_hold_queue);
1603
1604 mutex_lock(&knet->mutex);
1605 aggregate_mux_stats(&mux->stats, &knet->aggregate_mux_stats);
1606 aggregate_psock_stats(&mux->aggregate_psock_stats,
1607 &knet->aggregate_psock_stats);
1608 aggregate_strp_stats(&mux->aggregate_strp_stats,
1609 &knet->aggregate_strp_stats);
1610 list_del_rcu(&mux->kcm_mux_list);
1611 knet->count--;
1612 mutex_unlock(&knet->mutex);
1613
1614 kfree_rcu(mux, rcu);
1615}
1616
1617static void kcm_done(struct kcm_sock *kcm)
1618{
1619 struct kcm_mux *mux = kcm->mux;
1620 struct sock *sk = &kcm->sk;
1621 int socks_cnt;
1622
1623 spin_lock_bh(&mux->rx_lock);
1624 if (kcm->rx_psock) {
1625 /* Cleanup in unreserve_rx_kcm */
1626 WARN_ON(kcm->done);
1627 kcm->rx_disabled = 1;
1628 kcm->done = 1;
1629 spin_unlock_bh(&mux->rx_lock);
1630 return;
1631 }
1632
1633 if (kcm->rx_wait) {
1634 list_del(&kcm->wait_rx_list);
1635 /* paired with lockless reads in kcm_rfree() */
1636 WRITE_ONCE(kcm->rx_wait, false);
1637 }
1638 /* Move any pending receive messages to other kcm sockets */
1639 requeue_rx_msgs(mux, &sk->sk_receive_queue);
1640
1641 spin_unlock_bh(&mux->rx_lock);
1642
1643 if (WARN_ON(sk_rmem_alloc_get(sk)))
1644 return;
1645
1646 /* Detach from MUX */
1647 spin_lock_bh(&mux->lock);
1648
1649 list_del(&kcm->kcm_sock_list);
1650 mux->kcm_socks_cnt--;
1651 socks_cnt = mux->kcm_socks_cnt;
1652
1653 spin_unlock_bh(&mux->lock);
1654
1655 if (!socks_cnt) {
1656 /* We are done with the mux now. */
1657 release_mux(mux);
1658 }
1659
1660 WARN_ON(kcm->rx_wait);
1661
1662 sock_put(&kcm->sk);
1663}
1664
1665/* Called by kcm_release to close a KCM socket.
1666 * If this is the last KCM socket on the MUX, destroy the MUX.
1667 */
1668static int kcm_release(struct socket *sock)
1669{
1670 struct sock *sk = sock->sk;
1671 struct kcm_sock *kcm;
1672 struct kcm_mux *mux;
1673 struct kcm_psock *psock;
1674
1675 if (!sk)
1676 return 0;
1677
1678 kcm = kcm_sk(sk);
1679 mux = kcm->mux;
1680
1681 lock_sock(sk);
1682 sock_orphan(sk);
1683 kfree_skb(kcm->seq_skb);
1684
1685 /* Purge queue under lock to avoid race condition with tx_work trying
1686 * to act when queue is nonempty. If tx_work runs after this point
1687 * it will just return.
1688 */
1689 __skb_queue_purge(&sk->sk_write_queue);
1690
1691 /* Set tx_stopped. This is checked when psock is bound to a kcm and we
1692 * get a writespace callback. This prevents further work being queued
1693 * from the callback (unbinding the psock occurs after canceling work.
1694 */
1695 kcm->tx_stopped = 1;
1696
1697 release_sock(sk);
1698
1699 spin_lock_bh(&mux->lock);
1700 if (kcm->tx_wait) {
1701 /* Take of tx_wait list, after this point there should be no way
1702 * that a psock will be assigned to this kcm.
1703 */
1704 list_del(&kcm->wait_psock_list);
1705 kcm->tx_wait = false;
1706 }
1707 spin_unlock_bh(&mux->lock);
1708
1709 /* Cancel work. After this point there should be no outside references
1710 * to the kcm socket.
1711 */
1712 cancel_work_sync(&kcm->tx_work);
1713
1714 lock_sock(sk);
1715 psock = kcm->tx_psock;
1716 if (psock) {
1717 /* A psock was reserved, so we need to kill it since it
1718 * may already have some bytes queued from a message. We
1719 * need to do this after removing kcm from tx_wait list.
1720 */
1721 kcm_abort_tx_psock(psock, EPIPE, false);
1722 unreserve_psock(kcm);
1723 }
1724 release_sock(sk);
1725
1726 WARN_ON(kcm->tx_wait);
1727 WARN_ON(kcm->tx_psock);
1728
1729 sock->sk = NULL;
1730
1731 kcm_done(kcm);
1732
1733 return 0;
1734}
1735
1736static const struct proto_ops kcm_dgram_ops = {
1737 .family = PF_KCM,
1738 .owner = THIS_MODULE,
1739 .release = kcm_release,
1740 .bind = sock_no_bind,
1741 .connect = sock_no_connect,
1742 .socketpair = sock_no_socketpair,
1743 .accept = sock_no_accept,
1744 .getname = sock_no_getname,
1745 .poll = datagram_poll,
1746 .ioctl = kcm_ioctl,
1747 .listen = sock_no_listen,
1748 .shutdown = sock_no_shutdown,
1749 .setsockopt = kcm_setsockopt,
1750 .getsockopt = kcm_getsockopt,
1751 .sendmsg = kcm_sendmsg,
1752 .recvmsg = kcm_recvmsg,
1753 .mmap = sock_no_mmap,
1754 .splice_eof = kcm_splice_eof,
1755};
1756
1757static const struct proto_ops kcm_seqpacket_ops = {
1758 .family = PF_KCM,
1759 .owner = THIS_MODULE,
1760 .release = kcm_release,
1761 .bind = sock_no_bind,
1762 .connect = sock_no_connect,
1763 .socketpair = sock_no_socketpair,
1764 .accept = sock_no_accept,
1765 .getname = sock_no_getname,
1766 .poll = datagram_poll,
1767 .ioctl = kcm_ioctl,
1768 .listen = sock_no_listen,
1769 .shutdown = sock_no_shutdown,
1770 .setsockopt = kcm_setsockopt,
1771 .getsockopt = kcm_getsockopt,
1772 .sendmsg = kcm_sendmsg,
1773 .recvmsg = kcm_recvmsg,
1774 .mmap = sock_no_mmap,
1775 .splice_eof = kcm_splice_eof,
1776 .splice_read = kcm_splice_read,
1777};
1778
1779/* Create proto operation for kcm sockets */
1780static int kcm_create(struct net *net, struct socket *sock,
1781 int protocol, int kern)
1782{
1783 struct kcm_net *knet = net_generic(net, kcm_net_id);
1784 struct sock *sk;
1785 struct kcm_mux *mux;
1786
1787 switch (sock->type) {
1788 case SOCK_DGRAM:
1789 sock->ops = &kcm_dgram_ops;
1790 break;
1791 case SOCK_SEQPACKET:
1792 sock->ops = &kcm_seqpacket_ops;
1793 break;
1794 default:
1795 return -ESOCKTNOSUPPORT;
1796 }
1797
1798 if (protocol != KCMPROTO_CONNECTED)
1799 return -EPROTONOSUPPORT;
1800
1801 sk = sk_alloc(net, PF_KCM, GFP_KERNEL, &kcm_proto, kern);
1802 if (!sk)
1803 return -ENOMEM;
1804
1805 /* Allocate a kcm mux, shared between KCM sockets */
1806 mux = kmem_cache_zalloc(kcm_muxp, GFP_KERNEL);
1807 if (!mux) {
1808 sk_free(sk);
1809 return -ENOMEM;
1810 }
1811
1812 spin_lock_init(&mux->lock);
1813 spin_lock_init(&mux->rx_lock);
1814 INIT_LIST_HEAD(&mux->kcm_socks);
1815 INIT_LIST_HEAD(&mux->kcm_rx_waiters);
1816 INIT_LIST_HEAD(&mux->kcm_tx_waiters);
1817
1818 INIT_LIST_HEAD(&mux->psocks);
1819 INIT_LIST_HEAD(&mux->psocks_ready);
1820 INIT_LIST_HEAD(&mux->psocks_avail);
1821
1822 mux->knet = knet;
1823
1824 /* Add new MUX to list */
1825 mutex_lock(&knet->mutex);
1826 list_add_rcu(&mux->kcm_mux_list, &knet->mux_list);
1827 knet->count++;
1828 mutex_unlock(&knet->mutex);
1829
1830 skb_queue_head_init(&mux->rx_hold_queue);
1831
1832 /* Init KCM socket */
1833 sock_init_data(sock, sk);
1834 init_kcm_sock(kcm_sk(sk), mux);
1835
1836 return 0;
1837}
1838
1839static const struct net_proto_family kcm_family_ops = {
1840 .family = PF_KCM,
1841 .create = kcm_create,
1842 .owner = THIS_MODULE,
1843};
1844
1845static __net_init int kcm_init_net(struct net *net)
1846{
1847 struct kcm_net *knet = net_generic(net, kcm_net_id);
1848
1849 INIT_LIST_HEAD_RCU(&knet->mux_list);
1850 mutex_init(&knet->mutex);
1851
1852 return 0;
1853}
1854
1855static __net_exit void kcm_exit_net(struct net *net)
1856{
1857 struct kcm_net *knet = net_generic(net, kcm_net_id);
1858
1859 /* All KCM sockets should be closed at this point, which should mean
1860 * that all multiplexors and psocks have been destroyed.
1861 */
1862 WARN_ON(!list_empty(&knet->mux_list));
1863
1864 mutex_destroy(&knet->mutex);
1865}
1866
1867static struct pernet_operations kcm_net_ops = {
1868 .init = kcm_init_net,
1869 .exit = kcm_exit_net,
1870 .id = &kcm_net_id,
1871 .size = sizeof(struct kcm_net),
1872};
1873
1874static int __init kcm_init(void)
1875{
1876 int err = -ENOMEM;
1877
1878 kcm_muxp = KMEM_CACHE(kcm_mux, SLAB_HWCACHE_ALIGN);
1879 if (!kcm_muxp)
1880 goto fail;
1881
1882 kcm_psockp = KMEM_CACHE(kcm_psock, SLAB_HWCACHE_ALIGN);
1883 if (!kcm_psockp)
1884 goto fail;
1885
1886 kcm_wq = create_singlethread_workqueue("kkcmd");
1887 if (!kcm_wq)
1888 goto fail;
1889
1890 err = proto_register(&kcm_proto, 1);
1891 if (err)
1892 goto fail;
1893
1894 err = register_pernet_device(&kcm_net_ops);
1895 if (err)
1896 goto net_ops_fail;
1897
1898 err = sock_register(&kcm_family_ops);
1899 if (err)
1900 goto sock_register_fail;
1901
1902 err = kcm_proc_init();
1903 if (err)
1904 goto proc_init_fail;
1905
1906 return 0;
1907
1908proc_init_fail:
1909 sock_unregister(PF_KCM);
1910
1911sock_register_fail:
1912 unregister_pernet_device(&kcm_net_ops);
1913
1914net_ops_fail:
1915 proto_unregister(&kcm_proto);
1916
1917fail:
1918 kmem_cache_destroy(kcm_muxp);
1919 kmem_cache_destroy(kcm_psockp);
1920
1921 if (kcm_wq)
1922 destroy_workqueue(kcm_wq);
1923
1924 return err;
1925}
1926
1927static void __exit kcm_exit(void)
1928{
1929 kcm_proc_exit();
1930 sock_unregister(PF_KCM);
1931 unregister_pernet_device(&kcm_net_ops);
1932 proto_unregister(&kcm_proto);
1933 destroy_workqueue(kcm_wq);
1934
1935 kmem_cache_destroy(kcm_muxp);
1936 kmem_cache_destroy(kcm_psockp);
1937}
1938
1939module_init(kcm_init);
1940module_exit(kcm_exit);
1941
1942MODULE_LICENSE("GPL");
1943MODULE_DESCRIPTION("KCM (Kernel Connection Multiplexor) sockets");
1944MODULE_ALIAS_NETPROTO(PF_KCM);