xref: /linux/drivers/xen/pvcalls-back.c (revision 3d0fe49454652117522f60bfbefb978ba0e5300b)
1 // SPDX-License-Identifier: GPL-2.0-or-later
2 /*
3  * (c) 2017 Stefano Stabellini <stefano@aporeto.com>
4  */
5 
6 #include <linux/inet.h>
7 #include <linux/kthread.h>
8 #include <linux/list.h>
9 #include <linux/radix-tree.h>
10 #include <linux/module.h>
11 #include <linux/semaphore.h>
12 #include <linux/wait.h>
13 #include <net/sock.h>
14 #include <net/inet_common.h>
15 #include <net/inet_connection_sock.h>
16 #include <net/request_sock.h>
17 #include <trace/events/sock.h>
18 
19 #include <xen/events.h>
20 #include <xen/grant_table.h>
21 #include <xen/xen.h>
22 #include <xen/xenbus.h>
23 #include <xen/interface/io/pvcalls.h>
24 
25 #define PVCALLS_VERSIONS "1"
26 #define MAX_RING_ORDER XENBUS_MAX_RING_GRANT_ORDER
27 
28 static struct pvcalls_back_global {
29 	struct list_head frontends;
30 	struct semaphore frontends_lock;
31 } pvcalls_back_global;
32 
33 /*
34  * Per-frontend data structure. It contains pointers to the command
35  * ring, its event channel, a list of active sockets and a tree of
36  * passive sockets.
37  */
38 struct pvcalls_fedata {
39 	struct list_head list;
40 	struct xenbus_device *dev;
41 	struct xen_pvcalls_sring *sring;
42 	struct xen_pvcalls_back_ring ring;
43 	int irq;
44 	struct list_head socket_mappings;
45 	struct radix_tree_root socketpass_mappings;
46 	struct semaphore socket_lock;
47 };
48 
49 struct pvcalls_ioworker {
50 	struct work_struct register_work;
51 	struct workqueue_struct *wq;
52 };
53 
54 struct sock_mapping {
55 	struct list_head list;
56 	struct pvcalls_fedata *fedata;
57 	struct sockpass_mapping *sockpass;
58 	struct socket *sock;
59 	uint64_t id;
60 	grant_ref_t ref;
61 	struct pvcalls_data_intf *ring;
62 	void *bytes;
63 	struct pvcalls_data data;
64 	uint32_t ring_order;
65 	int irq;
66 	atomic_t read;
67 	atomic_t write;
68 	atomic_t io;
69 	atomic_t release;
70 	atomic_t eoi;
71 	void (*saved_data_ready)(struct sock *sk);
72 	struct pvcalls_ioworker ioworker;
73 };
74 
75 struct sockpass_mapping {
76 	struct list_head list;
77 	struct pvcalls_fedata *fedata;
78 	struct socket *sock;
79 	uint64_t id;
80 	struct xen_pvcalls_request reqcopy;
81 	spinlock_t copy_lock;
82 	struct workqueue_struct *wq;
83 	struct work_struct register_work;
84 	void (*saved_data_ready)(struct sock *sk);
85 };
86 
87 static irqreturn_t pvcalls_back_conn_event(int irq, void *sock_map);
88 static int pvcalls_back_release_active(struct xenbus_device *dev,
89 				       struct pvcalls_fedata *fedata,
90 				       struct sock_mapping *map);
91 
92 static bool pvcalls_conn_back_read(void *opaque)
93 {
94 	struct sock_mapping *map = (struct sock_mapping *)opaque;
95 	struct msghdr msg;
96 	struct kvec vec[2];
97 	RING_IDX cons, prod, size, wanted, array_size, masked_prod, masked_cons;
98 	int32_t error;
99 	struct pvcalls_data_intf *intf = map->ring;
100 	struct pvcalls_data *data = &map->data;
101 	unsigned long flags;
102 	int ret;
103 
104 	array_size = XEN_FLEX_RING_SIZE(map->ring_order);
105 	cons = intf->in_cons;
106 	prod = intf->in_prod;
107 	error = intf->in_error;
108 	/* read the indexes first, then deal with the data */
109 	virt_mb();
110 
111 	if (error)
112 		return false;
113 
114 	size = pvcalls_queued(prod, cons, array_size);
115 	if (size >= array_size)
116 		return false;
117 	spin_lock_irqsave(&map->sock->sk->sk_receive_queue.lock, flags);
118 	if (skb_queue_empty(&map->sock->sk->sk_receive_queue)) {
119 		atomic_set(&map->read, 0);
120 		spin_unlock_irqrestore(&map->sock->sk->sk_receive_queue.lock,
121 				flags);
122 		return true;
123 	}
124 	spin_unlock_irqrestore(&map->sock->sk->sk_receive_queue.lock, flags);
125 	wanted = array_size - size;
126 	masked_prod = pvcalls_mask(prod, array_size);
127 	masked_cons = pvcalls_mask(cons, array_size);
128 
129 	memset(&msg, 0, sizeof(msg));
130 	if (masked_prod < masked_cons) {
131 		vec[0].iov_base = data->in + masked_prod;
132 		vec[0].iov_len = wanted;
133 		iov_iter_kvec(&msg.msg_iter, ITER_DEST, vec, 1, wanted);
134 	} else {
135 		vec[0].iov_base = data->in + masked_prod;
136 		vec[0].iov_len = array_size - masked_prod;
137 		vec[1].iov_base = data->in;
138 		vec[1].iov_len = wanted - vec[0].iov_len;
139 		iov_iter_kvec(&msg.msg_iter, ITER_DEST, vec, 2, wanted);
140 	}
141 
142 	atomic_set(&map->read, 0);
143 	ret = inet_recvmsg(map->sock, &msg, wanted, MSG_DONTWAIT);
144 	WARN_ON(ret > wanted);
145 	if (ret == -EAGAIN) /* shouldn't happen */
146 		return true;
147 	if (!ret)
148 		ret = -ENOTCONN;
149 	spin_lock_irqsave(&map->sock->sk->sk_receive_queue.lock, flags);
150 	if (ret > 0 && !skb_queue_empty(&map->sock->sk->sk_receive_queue))
151 		atomic_inc(&map->read);
152 	spin_unlock_irqrestore(&map->sock->sk->sk_receive_queue.lock, flags);
153 
154 	/* write the data, then modify the indexes */
155 	virt_wmb();
156 	if (ret < 0) {
157 		atomic_set(&map->read, 0);
158 		intf->in_error = ret;
159 	} else
160 		intf->in_prod = prod + ret;
161 	/* update the indexes, then notify the other end */
162 	virt_wmb();
163 	notify_remote_via_irq(map->irq);
164 
165 	return true;
166 }
167 
168 static bool pvcalls_conn_back_write(struct sock_mapping *map)
169 {
170 	struct pvcalls_data_intf *intf = map->ring;
171 	struct pvcalls_data *data = &map->data;
172 	struct msghdr msg;
173 	struct kvec vec[2];
174 	RING_IDX cons, prod, size, array_size;
175 	int ret;
176 
177 	atomic_set(&map->write, 0);
178 
179 	cons = intf->out_cons;
180 	prod = intf->out_prod;
181 	/* read the indexes before dealing with the data */
182 	virt_mb();
183 
184 	array_size = XEN_FLEX_RING_SIZE(map->ring_order);
185 	size = pvcalls_queued(prod, cons, array_size);
186 	if (size == 0)
187 		return false;
188 
189 	memset(&msg, 0, sizeof(msg));
190 	msg.msg_flags |= MSG_DONTWAIT;
191 	if (pvcalls_mask(prod, array_size) > pvcalls_mask(cons, array_size)) {
192 		vec[0].iov_base = data->out + pvcalls_mask(cons, array_size);
193 		vec[0].iov_len = size;
194 		iov_iter_kvec(&msg.msg_iter, ITER_SOURCE, vec, 1, size);
195 	} else {
196 		vec[0].iov_base = data->out + pvcalls_mask(cons, array_size);
197 		vec[0].iov_len = array_size - pvcalls_mask(cons, array_size);
198 		vec[1].iov_base = data->out;
199 		vec[1].iov_len = size - vec[0].iov_len;
200 		iov_iter_kvec(&msg.msg_iter, ITER_SOURCE, vec, 2, size);
201 	}
202 
203 	ret = inet_sendmsg(map->sock, &msg, size);
204 	if (ret == -EAGAIN) {
205 		atomic_inc(&map->write);
206 		atomic_inc(&map->io);
207 		return true;
208 	}
209 
210 	/* write the data, then update the indexes */
211 	virt_wmb();
212 	if (ret < 0) {
213 		intf->out_error = ret;
214 	} else {
215 		intf->out_error = 0;
216 		intf->out_cons = cons + ret;
217 		prod = intf->out_prod;
218 	}
219 	/* update the indexes, then notify the other end */
220 	virt_wmb();
221 	if (prod != cons + ret) {
222 		atomic_inc(&map->write);
223 		atomic_inc(&map->io);
224 	}
225 	notify_remote_via_irq(map->irq);
226 
227 	return true;
228 }
229 
230 static void pvcalls_back_ioworker(struct work_struct *work)
231 {
232 	struct pvcalls_ioworker *ioworker = container_of(work,
233 		struct pvcalls_ioworker, register_work);
234 	struct sock_mapping *map = container_of(ioworker, struct sock_mapping,
235 		ioworker);
236 	unsigned int eoi_flags = XEN_EOI_FLAG_SPURIOUS;
237 
238 	while (atomic_read(&map->io) > 0) {
239 		if (atomic_read(&map->release) > 0) {
240 			atomic_set(&map->release, 0);
241 			return;
242 		}
243 
244 		if (atomic_read(&map->read) > 0 &&
245 		    pvcalls_conn_back_read(map))
246 			eoi_flags = 0;
247 		if (atomic_read(&map->write) > 0 &&
248 		    pvcalls_conn_back_write(map))
249 			eoi_flags = 0;
250 
251 		if (atomic_read(&map->eoi) > 0 && !atomic_read(&map->write)) {
252 			atomic_set(&map->eoi, 0);
253 			xen_irq_lateeoi(map->irq, eoi_flags);
254 			eoi_flags = XEN_EOI_FLAG_SPURIOUS;
255 		}
256 
257 		atomic_dec(&map->io);
258 	}
259 }
260 
261 static int pvcalls_back_socket(struct xenbus_device *dev,
262 		struct xen_pvcalls_request *req)
263 {
264 	struct pvcalls_fedata *fedata;
265 	int ret;
266 	struct xen_pvcalls_response *rsp;
267 
268 	fedata = dev_get_drvdata(&dev->dev);
269 
270 	if (req->u.socket.domain != AF_INET ||
271 	    req->u.socket.type != SOCK_STREAM ||
272 	    (req->u.socket.protocol != IPPROTO_IP &&
273 	     req->u.socket.protocol != AF_INET))
274 		ret = -EAFNOSUPPORT;
275 	else
276 		ret = 0;
277 
278 	/* leave the actual socket allocation for later */
279 
280 	rsp = RING_GET_RESPONSE(&fedata->ring, fedata->ring.rsp_prod_pvt++);
281 	rsp->req_id = req->req_id;
282 	rsp->cmd = req->cmd;
283 	rsp->u.socket.id = req->u.socket.id;
284 	rsp->ret = ret;
285 
286 	return 0;
287 }
288 
289 static void pvcalls_sk_state_change(struct sock *sock)
290 {
291 	struct sock_mapping *map = sock->sk_user_data;
292 
293 	if (map == NULL)
294 		return;
295 
296 	atomic_inc(&map->read);
297 	notify_remote_via_irq(map->irq);
298 }
299 
300 static void pvcalls_sk_data_ready(struct sock *sock)
301 {
302 	struct sock_mapping *map = sock->sk_user_data;
303 	struct pvcalls_ioworker *iow;
304 
305 	trace_sk_data_ready(sock);
306 
307 	if (map == NULL)
308 		return;
309 
310 	iow = &map->ioworker;
311 	atomic_inc(&map->read);
312 	atomic_inc(&map->io);
313 	queue_work(iow->wq, &iow->register_work);
314 }
315 
316 static struct sock_mapping *pvcalls_new_active_socket(
317 		struct pvcalls_fedata *fedata,
318 		uint64_t id,
319 		grant_ref_t ref,
320 		evtchn_port_t evtchn,
321 		struct socket *sock)
322 {
323 	int ret;
324 	struct sock_mapping *map;
325 	void *page;
326 
327 	map = kzalloc(sizeof(*map), GFP_KERNEL);
328 	if (map == NULL) {
329 		sock_release(sock);
330 		return NULL;
331 	}
332 
333 	map->fedata = fedata;
334 	map->sock = sock;
335 	map->id = id;
336 	map->ref = ref;
337 
338 	ret = xenbus_map_ring_valloc(fedata->dev, &ref, 1, &page);
339 	if (ret < 0)
340 		goto out;
341 	map->ring = page;
342 	map->ring_order = map->ring->ring_order;
343 	/* first read the order, then map the data ring */
344 	virt_rmb();
345 	if (map->ring_order > MAX_RING_ORDER) {
346 		pr_warn("%s frontend requested ring_order %u, which is > MAX (%u)\n",
347 				__func__, map->ring_order, MAX_RING_ORDER);
348 		goto out;
349 	}
350 	ret = xenbus_map_ring_valloc(fedata->dev, map->ring->ref,
351 				     (1 << map->ring_order), &page);
352 	if (ret < 0)
353 		goto out;
354 	map->bytes = page;
355 
356 	ret = bind_interdomain_evtchn_to_irqhandler_lateeoi(
357 			fedata->dev, evtchn,
358 			pvcalls_back_conn_event, 0, "pvcalls-backend", map);
359 	if (ret < 0)
360 		goto out;
361 	map->irq = ret;
362 
363 	map->data.in = map->bytes;
364 	map->data.out = map->bytes + XEN_FLEX_RING_SIZE(map->ring_order);
365 
366 	map->ioworker.wq = alloc_ordered_workqueue("pvcalls_io", 0);
367 	if (!map->ioworker.wq)
368 		goto out;
369 	atomic_set(&map->io, 1);
370 	INIT_WORK(&map->ioworker.register_work,	pvcalls_back_ioworker);
371 
372 	down(&fedata->socket_lock);
373 	list_add_tail(&map->list, &fedata->socket_mappings);
374 	up(&fedata->socket_lock);
375 
376 	write_lock_bh(&map->sock->sk->sk_callback_lock);
377 	map->saved_data_ready = map->sock->sk->sk_data_ready;
378 	map->sock->sk->sk_user_data = map;
379 	map->sock->sk->sk_data_ready = pvcalls_sk_data_ready;
380 	map->sock->sk->sk_state_change = pvcalls_sk_state_change;
381 	write_unlock_bh(&map->sock->sk->sk_callback_lock);
382 
383 	return map;
384 out:
385 	down(&fedata->socket_lock);
386 	list_del(&map->list);
387 	pvcalls_back_release_active(fedata->dev, fedata, map);
388 	up(&fedata->socket_lock);
389 	return NULL;
390 }
391 
392 static int pvcalls_back_connect(struct xenbus_device *dev,
393 				struct xen_pvcalls_request *req)
394 {
395 	struct pvcalls_fedata *fedata;
396 	int ret = -EINVAL;
397 	struct socket *sock;
398 	struct sock_mapping *map;
399 	struct xen_pvcalls_response *rsp;
400 	struct sockaddr *sa = (struct sockaddr *)&req->u.connect.addr;
401 
402 	fedata = dev_get_drvdata(&dev->dev);
403 
404 	if (req->u.connect.len < sizeof(sa->sa_family) ||
405 	    req->u.connect.len > sizeof(req->u.connect.addr) ||
406 	    sa->sa_family != AF_INET)
407 		goto out;
408 
409 	ret = sock_create(AF_INET, SOCK_STREAM, 0, &sock);
410 	if (ret < 0)
411 		goto out;
412 	ret = inet_stream_connect(sock, sa, req->u.connect.len, 0);
413 	if (ret < 0) {
414 		sock_release(sock);
415 		goto out;
416 	}
417 
418 	map = pvcalls_new_active_socket(fedata,
419 					req->u.connect.id,
420 					req->u.connect.ref,
421 					req->u.connect.evtchn,
422 					sock);
423 	if (!map)
424 		ret = -EFAULT;
425 
426 out:
427 	rsp = RING_GET_RESPONSE(&fedata->ring, fedata->ring.rsp_prod_pvt++);
428 	rsp->req_id = req->req_id;
429 	rsp->cmd = req->cmd;
430 	rsp->u.connect.id = req->u.connect.id;
431 	rsp->ret = ret;
432 
433 	return 0;
434 }
435 
436 static int pvcalls_back_release_active(struct xenbus_device *dev,
437 				       struct pvcalls_fedata *fedata,
438 				       struct sock_mapping *map)
439 {
440 	disable_irq(map->irq);
441 	if (map->sock->sk != NULL) {
442 		write_lock_bh(&map->sock->sk->sk_callback_lock);
443 		map->sock->sk->sk_user_data = NULL;
444 		map->sock->sk->sk_data_ready = map->saved_data_ready;
445 		write_unlock_bh(&map->sock->sk->sk_callback_lock);
446 	}
447 
448 	atomic_set(&map->release, 1);
449 	flush_work(&map->ioworker.register_work);
450 
451 	xenbus_unmap_ring_vfree(dev, map->bytes);
452 	xenbus_unmap_ring_vfree(dev, (void *)map->ring);
453 	unbind_from_irqhandler(map->irq, map);
454 
455 	sock_release(map->sock);
456 	kfree(map);
457 
458 	return 0;
459 }
460 
461 static int pvcalls_back_release_passive(struct xenbus_device *dev,
462 					struct pvcalls_fedata *fedata,
463 					struct sockpass_mapping *mappass)
464 {
465 	if (mappass->sock->sk != NULL) {
466 		write_lock_bh(&mappass->sock->sk->sk_callback_lock);
467 		mappass->sock->sk->sk_user_data = NULL;
468 		mappass->sock->sk->sk_data_ready = mappass->saved_data_ready;
469 		write_unlock_bh(&mappass->sock->sk->sk_callback_lock);
470 	}
471 	sock_release(mappass->sock);
472 	destroy_workqueue(mappass->wq);
473 	kfree(mappass);
474 
475 	return 0;
476 }
477 
478 static int pvcalls_back_release(struct xenbus_device *dev,
479 				struct xen_pvcalls_request *req)
480 {
481 	struct pvcalls_fedata *fedata;
482 	struct sock_mapping *map, *n;
483 	struct sockpass_mapping *mappass;
484 	int ret = 0;
485 	struct xen_pvcalls_response *rsp;
486 
487 	fedata = dev_get_drvdata(&dev->dev);
488 
489 	down(&fedata->socket_lock);
490 	list_for_each_entry_safe(map, n, &fedata->socket_mappings, list) {
491 		if (map->id == req->u.release.id) {
492 			list_del(&map->list);
493 			up(&fedata->socket_lock);
494 			ret = pvcalls_back_release_active(dev, fedata, map);
495 			goto out;
496 		}
497 	}
498 	mappass = radix_tree_lookup(&fedata->socketpass_mappings,
499 				    req->u.release.id);
500 	if (mappass != NULL) {
501 		radix_tree_delete(&fedata->socketpass_mappings, mappass->id);
502 		up(&fedata->socket_lock);
503 		ret = pvcalls_back_release_passive(dev, fedata, mappass);
504 	} else
505 		up(&fedata->socket_lock);
506 
507 out:
508 	rsp = RING_GET_RESPONSE(&fedata->ring, fedata->ring.rsp_prod_pvt++);
509 	rsp->req_id = req->req_id;
510 	rsp->u.release.id = req->u.release.id;
511 	rsp->cmd = req->cmd;
512 	rsp->ret = ret;
513 	return 0;
514 }
515 
516 static void __pvcalls_back_accept(struct work_struct *work)
517 {
518 	struct sockpass_mapping *mappass = container_of(
519 		work, struct sockpass_mapping, register_work);
520 	struct sock_mapping *map;
521 	struct pvcalls_ioworker *iow;
522 	struct pvcalls_fedata *fedata;
523 	struct socket *sock;
524 	struct xen_pvcalls_response *rsp;
525 	struct xen_pvcalls_request *req;
526 	int notify;
527 	int ret = -EINVAL;
528 	unsigned long flags;
529 
530 	fedata = mappass->fedata;
531 	/*
532 	 * __pvcalls_back_accept can race against pvcalls_back_accept.
533 	 * We only need to check the value of "cmd" on read. It could be
534 	 * done atomically, but to simplify the code on the write side, we
535 	 * use a spinlock.
536 	 */
537 	spin_lock_irqsave(&mappass->copy_lock, flags);
538 	req = &mappass->reqcopy;
539 	if (req->cmd != PVCALLS_ACCEPT) {
540 		spin_unlock_irqrestore(&mappass->copy_lock, flags);
541 		return;
542 	}
543 	spin_unlock_irqrestore(&mappass->copy_lock, flags);
544 
545 	sock = sock_alloc();
546 	if (sock == NULL)
547 		goto out_error;
548 	sock->type = mappass->sock->type;
549 	sock->ops = mappass->sock->ops;
550 
551 	ret = inet_accept(mappass->sock, sock, O_NONBLOCK, true);
552 	if (ret == -EAGAIN) {
553 		sock_release(sock);
554 		return;
555 	}
556 
557 	map = pvcalls_new_active_socket(fedata,
558 					req->u.accept.id_new,
559 					req->u.accept.ref,
560 					req->u.accept.evtchn,
561 					sock);
562 	if (!map) {
563 		ret = -EFAULT;
564 		goto out_error;
565 	}
566 
567 	map->sockpass = mappass;
568 	iow = &map->ioworker;
569 	atomic_inc(&map->read);
570 	atomic_inc(&map->io);
571 	queue_work(iow->wq, &iow->register_work);
572 
573 out_error:
574 	rsp = RING_GET_RESPONSE(&fedata->ring, fedata->ring.rsp_prod_pvt++);
575 	rsp->req_id = req->req_id;
576 	rsp->cmd = req->cmd;
577 	rsp->u.accept.id = req->u.accept.id;
578 	rsp->ret = ret;
579 	RING_PUSH_RESPONSES_AND_CHECK_NOTIFY(&fedata->ring, notify);
580 	if (notify)
581 		notify_remote_via_irq(fedata->irq);
582 
583 	mappass->reqcopy.cmd = 0;
584 }
585 
586 static void pvcalls_pass_sk_data_ready(struct sock *sock)
587 {
588 	struct sockpass_mapping *mappass = sock->sk_user_data;
589 	struct pvcalls_fedata *fedata;
590 	struct xen_pvcalls_response *rsp;
591 	unsigned long flags;
592 	int notify;
593 
594 	trace_sk_data_ready(sock);
595 
596 	if (mappass == NULL)
597 		return;
598 
599 	fedata = mappass->fedata;
600 	spin_lock_irqsave(&mappass->copy_lock, flags);
601 	if (mappass->reqcopy.cmd == PVCALLS_POLL) {
602 		rsp = RING_GET_RESPONSE(&fedata->ring,
603 					fedata->ring.rsp_prod_pvt++);
604 		rsp->req_id = mappass->reqcopy.req_id;
605 		rsp->u.poll.id = mappass->reqcopy.u.poll.id;
606 		rsp->cmd = mappass->reqcopy.cmd;
607 		rsp->ret = 0;
608 
609 		mappass->reqcopy.cmd = 0;
610 		spin_unlock_irqrestore(&mappass->copy_lock, flags);
611 
612 		RING_PUSH_RESPONSES_AND_CHECK_NOTIFY(&fedata->ring, notify);
613 		if (notify)
614 			notify_remote_via_irq(mappass->fedata->irq);
615 	} else {
616 		spin_unlock_irqrestore(&mappass->copy_lock, flags);
617 		queue_work(mappass->wq, &mappass->register_work);
618 	}
619 }
620 
621 static int pvcalls_back_bind(struct xenbus_device *dev,
622 			     struct xen_pvcalls_request *req)
623 {
624 	struct pvcalls_fedata *fedata;
625 	int ret;
626 	struct sockpass_mapping *map;
627 	struct xen_pvcalls_response *rsp;
628 
629 	fedata = dev_get_drvdata(&dev->dev);
630 
631 	map = kzalloc(sizeof(*map), GFP_KERNEL);
632 	if (map == NULL) {
633 		ret = -ENOMEM;
634 		goto out;
635 	}
636 
637 	INIT_WORK(&map->register_work, __pvcalls_back_accept);
638 	spin_lock_init(&map->copy_lock);
639 	map->wq = alloc_ordered_workqueue("pvcalls_wq", 0);
640 	if (!map->wq) {
641 		ret = -ENOMEM;
642 		goto out;
643 	}
644 
645 	ret = sock_create(AF_INET, SOCK_STREAM, 0, &map->sock);
646 	if (ret < 0)
647 		goto out;
648 
649 	ret = inet_bind(map->sock, (struct sockaddr *)&req->u.bind.addr,
650 			req->u.bind.len);
651 	if (ret < 0)
652 		goto out;
653 
654 	map->fedata = fedata;
655 	map->id = req->u.bind.id;
656 
657 	down(&fedata->socket_lock);
658 	ret = radix_tree_insert(&fedata->socketpass_mappings, map->id,
659 				map);
660 	up(&fedata->socket_lock);
661 	if (ret)
662 		goto out;
663 
664 	write_lock_bh(&map->sock->sk->sk_callback_lock);
665 	map->saved_data_ready = map->sock->sk->sk_data_ready;
666 	map->sock->sk->sk_user_data = map;
667 	map->sock->sk->sk_data_ready = pvcalls_pass_sk_data_ready;
668 	write_unlock_bh(&map->sock->sk->sk_callback_lock);
669 
670 out:
671 	if (ret) {
672 		if (map && map->sock)
673 			sock_release(map->sock);
674 		if (map && map->wq)
675 			destroy_workqueue(map->wq);
676 		kfree(map);
677 	}
678 	rsp = RING_GET_RESPONSE(&fedata->ring, fedata->ring.rsp_prod_pvt++);
679 	rsp->req_id = req->req_id;
680 	rsp->cmd = req->cmd;
681 	rsp->u.bind.id = req->u.bind.id;
682 	rsp->ret = ret;
683 	return 0;
684 }
685 
686 static int pvcalls_back_listen(struct xenbus_device *dev,
687 			       struct xen_pvcalls_request *req)
688 {
689 	struct pvcalls_fedata *fedata;
690 	int ret = -EINVAL;
691 	struct sockpass_mapping *map;
692 	struct xen_pvcalls_response *rsp;
693 
694 	fedata = dev_get_drvdata(&dev->dev);
695 
696 	down(&fedata->socket_lock);
697 	map = radix_tree_lookup(&fedata->socketpass_mappings, req->u.listen.id);
698 	up(&fedata->socket_lock);
699 	if (map == NULL)
700 		goto out;
701 
702 	ret = inet_listen(map->sock, req->u.listen.backlog);
703 
704 out:
705 	rsp = RING_GET_RESPONSE(&fedata->ring, fedata->ring.rsp_prod_pvt++);
706 	rsp->req_id = req->req_id;
707 	rsp->cmd = req->cmd;
708 	rsp->u.listen.id = req->u.listen.id;
709 	rsp->ret = ret;
710 	return 0;
711 }
712 
713 static int pvcalls_back_accept(struct xenbus_device *dev,
714 			       struct xen_pvcalls_request *req)
715 {
716 	struct pvcalls_fedata *fedata;
717 	struct sockpass_mapping *mappass;
718 	int ret = -EINVAL;
719 	struct xen_pvcalls_response *rsp;
720 	unsigned long flags;
721 
722 	fedata = dev_get_drvdata(&dev->dev);
723 
724 	down(&fedata->socket_lock);
725 	mappass = radix_tree_lookup(&fedata->socketpass_mappings,
726 		req->u.accept.id);
727 	up(&fedata->socket_lock);
728 	if (mappass == NULL)
729 		goto out_error;
730 
731 	/*
732 	 * Limitation of the current implementation: only support one
733 	 * concurrent accept or poll call on one socket.
734 	 */
735 	spin_lock_irqsave(&mappass->copy_lock, flags);
736 	if (mappass->reqcopy.cmd != 0) {
737 		spin_unlock_irqrestore(&mappass->copy_lock, flags);
738 		ret = -EINTR;
739 		goto out_error;
740 	}
741 
742 	mappass->reqcopy = *req;
743 	spin_unlock_irqrestore(&mappass->copy_lock, flags);
744 	queue_work(mappass->wq, &mappass->register_work);
745 
746 	/* Tell the caller we don't need to send back a notification yet */
747 	return -1;
748 
749 out_error:
750 	rsp = RING_GET_RESPONSE(&fedata->ring, fedata->ring.rsp_prod_pvt++);
751 	rsp->req_id = req->req_id;
752 	rsp->cmd = req->cmd;
753 	rsp->u.accept.id = req->u.accept.id;
754 	rsp->ret = ret;
755 	return 0;
756 }
757 
758 static int pvcalls_back_poll(struct xenbus_device *dev,
759 			     struct xen_pvcalls_request *req)
760 {
761 	struct pvcalls_fedata *fedata;
762 	struct sockpass_mapping *mappass;
763 	struct xen_pvcalls_response *rsp;
764 	struct inet_connection_sock *icsk;
765 	struct request_sock_queue *queue;
766 	unsigned long flags;
767 	int ret;
768 	bool data;
769 
770 	fedata = dev_get_drvdata(&dev->dev);
771 
772 	down(&fedata->socket_lock);
773 	mappass = radix_tree_lookup(&fedata->socketpass_mappings,
774 				    req->u.poll.id);
775 	up(&fedata->socket_lock);
776 	if (mappass == NULL)
777 		return -EINVAL;
778 
779 	/*
780 	 * Limitation of the current implementation: only support one
781 	 * concurrent accept or poll call on one socket.
782 	 */
783 	spin_lock_irqsave(&mappass->copy_lock, flags);
784 	if (mappass->reqcopy.cmd != 0) {
785 		ret = -EINTR;
786 		goto out;
787 	}
788 
789 	mappass->reqcopy = *req;
790 	icsk = inet_csk(mappass->sock->sk);
791 	queue = &icsk->icsk_accept_queue;
792 	data = READ_ONCE(queue->rskq_accept_head) != NULL;
793 	if (data) {
794 		mappass->reqcopy.cmd = 0;
795 		ret = 0;
796 		goto out;
797 	}
798 	spin_unlock_irqrestore(&mappass->copy_lock, flags);
799 
800 	/* Tell the caller we don't need to send back a notification yet */
801 	return -1;
802 
803 out:
804 	spin_unlock_irqrestore(&mappass->copy_lock, flags);
805 
806 	rsp = RING_GET_RESPONSE(&fedata->ring, fedata->ring.rsp_prod_pvt++);
807 	rsp->req_id = req->req_id;
808 	rsp->cmd = req->cmd;
809 	rsp->u.poll.id = req->u.poll.id;
810 	rsp->ret = ret;
811 	return 0;
812 }
813 
814 static int pvcalls_back_handle_cmd(struct xenbus_device *dev,
815 				   struct xen_pvcalls_request *req)
816 {
817 	int ret = 0;
818 
819 	switch (req->cmd) {
820 	case PVCALLS_SOCKET:
821 		ret = pvcalls_back_socket(dev, req);
822 		break;
823 	case PVCALLS_CONNECT:
824 		ret = pvcalls_back_connect(dev, req);
825 		break;
826 	case PVCALLS_RELEASE:
827 		ret = pvcalls_back_release(dev, req);
828 		break;
829 	case PVCALLS_BIND:
830 		ret = pvcalls_back_bind(dev, req);
831 		break;
832 	case PVCALLS_LISTEN:
833 		ret = pvcalls_back_listen(dev, req);
834 		break;
835 	case PVCALLS_ACCEPT:
836 		ret = pvcalls_back_accept(dev, req);
837 		break;
838 	case PVCALLS_POLL:
839 		ret = pvcalls_back_poll(dev, req);
840 		break;
841 	default:
842 	{
843 		struct pvcalls_fedata *fedata;
844 		struct xen_pvcalls_response *rsp;
845 
846 		fedata = dev_get_drvdata(&dev->dev);
847 		rsp = RING_GET_RESPONSE(
848 				&fedata->ring, fedata->ring.rsp_prod_pvt++);
849 		rsp->req_id = req->req_id;
850 		rsp->cmd = req->cmd;
851 		rsp->ret = -ENOTSUPP;
852 		break;
853 	}
854 	}
855 	return ret;
856 }
857 
858 static void pvcalls_back_work(struct pvcalls_fedata *fedata)
859 {
860 	int notify, notify_all = 0, more = 1;
861 	struct xen_pvcalls_request req;
862 	struct xenbus_device *dev = fedata->dev;
863 
864 	while (more) {
865 		while (RING_HAS_UNCONSUMED_REQUESTS(&fedata->ring)) {
866 			RING_COPY_REQUEST(&fedata->ring,
867 					  fedata->ring.req_cons++,
868 					  &req);
869 
870 			if (!pvcalls_back_handle_cmd(dev, &req)) {
871 				RING_PUSH_RESPONSES_AND_CHECK_NOTIFY(
872 					&fedata->ring, notify);
873 				notify_all += notify;
874 			}
875 		}
876 
877 		if (notify_all) {
878 			notify_remote_via_irq(fedata->irq);
879 			notify_all = 0;
880 		}
881 
882 		RING_FINAL_CHECK_FOR_REQUESTS(&fedata->ring, more);
883 	}
884 }
885 
886 static irqreturn_t pvcalls_back_event(int irq, void *dev_id)
887 {
888 	struct xenbus_device *dev = dev_id;
889 	struct pvcalls_fedata *fedata = NULL;
890 	unsigned int eoi_flags = XEN_EOI_FLAG_SPURIOUS;
891 
892 	if (dev) {
893 		fedata = dev_get_drvdata(&dev->dev);
894 		if (fedata) {
895 			pvcalls_back_work(fedata);
896 			eoi_flags = 0;
897 		}
898 	}
899 
900 	xen_irq_lateeoi(irq, eoi_flags);
901 
902 	return IRQ_HANDLED;
903 }
904 
905 static irqreturn_t pvcalls_back_conn_event(int irq, void *sock_map)
906 {
907 	struct sock_mapping *map = sock_map;
908 	struct pvcalls_ioworker *iow;
909 
910 	if (map == NULL || map->sock == NULL || map->sock->sk == NULL ||
911 		map->sock->sk->sk_user_data != map) {
912 		xen_irq_lateeoi(irq, 0);
913 		return IRQ_HANDLED;
914 	}
915 
916 	iow = &map->ioworker;
917 
918 	atomic_inc(&map->write);
919 	atomic_inc(&map->eoi);
920 	atomic_inc(&map->io);
921 	queue_work(iow->wq, &iow->register_work);
922 
923 	return IRQ_HANDLED;
924 }
925 
926 static int backend_connect(struct xenbus_device *dev)
927 {
928 	int err;
929 	evtchn_port_t evtchn;
930 	grant_ref_t ring_ref;
931 	struct pvcalls_fedata *fedata = NULL;
932 
933 	fedata = kzalloc(sizeof(struct pvcalls_fedata), GFP_KERNEL);
934 	if (!fedata)
935 		return -ENOMEM;
936 
937 	fedata->irq = -1;
938 	err = xenbus_scanf(XBT_NIL, dev->otherend, "port", "%u",
939 			   &evtchn);
940 	if (err != 1) {
941 		err = -EINVAL;
942 		xenbus_dev_fatal(dev, err, "reading %s/event-channel",
943 				 dev->otherend);
944 		goto error;
945 	}
946 
947 	err = xenbus_scanf(XBT_NIL, dev->otherend, "ring-ref", "%u", &ring_ref);
948 	if (err != 1) {
949 		err = -EINVAL;
950 		xenbus_dev_fatal(dev, err, "reading %s/ring-ref",
951 				 dev->otherend);
952 		goto error;
953 	}
954 
955 	err = bind_interdomain_evtchn_to_irq_lateeoi(dev, evtchn);
956 	if (err < 0)
957 		goto error;
958 	fedata->irq = err;
959 
960 	err = request_threaded_irq(fedata->irq, NULL, pvcalls_back_event,
961 				   IRQF_ONESHOT, "pvcalls-back", dev);
962 	if (err < 0)
963 		goto error;
964 
965 	err = xenbus_map_ring_valloc(dev, &ring_ref, 1,
966 				     (void **)&fedata->sring);
967 	if (err < 0)
968 		goto error;
969 
970 	BACK_RING_INIT(&fedata->ring, fedata->sring, XEN_PAGE_SIZE * 1);
971 	fedata->dev = dev;
972 
973 	INIT_LIST_HEAD(&fedata->socket_mappings);
974 	INIT_RADIX_TREE(&fedata->socketpass_mappings, GFP_KERNEL);
975 	sema_init(&fedata->socket_lock, 1);
976 	dev_set_drvdata(&dev->dev, fedata);
977 
978 	down(&pvcalls_back_global.frontends_lock);
979 	list_add_tail(&fedata->list, &pvcalls_back_global.frontends);
980 	up(&pvcalls_back_global.frontends_lock);
981 
982 	return 0;
983 
984  error:
985 	if (fedata->irq >= 0)
986 		unbind_from_irqhandler(fedata->irq, dev);
987 	if (fedata->sring != NULL)
988 		xenbus_unmap_ring_vfree(dev, fedata->sring);
989 	kfree(fedata);
990 	return err;
991 }
992 
993 static int backend_disconnect(struct xenbus_device *dev)
994 {
995 	struct pvcalls_fedata *fedata;
996 	struct sock_mapping *map, *n;
997 	struct sockpass_mapping *mappass;
998 	struct radix_tree_iter iter;
999 	void **slot;
1000 
1001 
1002 	fedata = dev_get_drvdata(&dev->dev);
1003 
1004 	down(&fedata->socket_lock);
1005 	list_for_each_entry_safe(map, n, &fedata->socket_mappings, list) {
1006 		list_del(&map->list);
1007 		pvcalls_back_release_active(dev, fedata, map);
1008 	}
1009 
1010 	radix_tree_for_each_slot(slot, &fedata->socketpass_mappings, &iter, 0) {
1011 		mappass = radix_tree_deref_slot(slot);
1012 		if (!mappass)
1013 			continue;
1014 		if (radix_tree_exception(mappass)) {
1015 			if (radix_tree_deref_retry(mappass))
1016 				slot = radix_tree_iter_retry(&iter);
1017 		} else {
1018 			radix_tree_delete(&fedata->socketpass_mappings,
1019 					  mappass->id);
1020 			pvcalls_back_release_passive(dev, fedata, mappass);
1021 		}
1022 	}
1023 	up(&fedata->socket_lock);
1024 
1025 	unbind_from_irqhandler(fedata->irq, dev);
1026 	xenbus_unmap_ring_vfree(dev, fedata->sring);
1027 
1028 	list_del(&fedata->list);
1029 	kfree(fedata);
1030 	dev_set_drvdata(&dev->dev, NULL);
1031 
1032 	return 0;
1033 }
1034 
1035 static int pvcalls_back_probe(struct xenbus_device *dev,
1036 			      const struct xenbus_device_id *id)
1037 {
1038 	int err, abort;
1039 	struct xenbus_transaction xbt;
1040 
1041 again:
1042 	abort = 1;
1043 
1044 	err = xenbus_transaction_start(&xbt);
1045 	if (err) {
1046 		pr_warn("%s cannot create xenstore transaction\n", __func__);
1047 		return err;
1048 	}
1049 
1050 	err = xenbus_printf(xbt, dev->nodename, "versions", "%s",
1051 			    PVCALLS_VERSIONS);
1052 	if (err) {
1053 		pr_warn("%s write out 'versions' failed\n", __func__);
1054 		goto abort;
1055 	}
1056 
1057 	err = xenbus_printf(xbt, dev->nodename, "max-page-order", "%u",
1058 			    MAX_RING_ORDER);
1059 	if (err) {
1060 		pr_warn("%s write out 'max-page-order' failed\n", __func__);
1061 		goto abort;
1062 	}
1063 
1064 	err = xenbus_printf(xbt, dev->nodename, "function-calls",
1065 			    XENBUS_FUNCTIONS_CALLS);
1066 	if (err) {
1067 		pr_warn("%s write out 'function-calls' failed\n", __func__);
1068 		goto abort;
1069 	}
1070 
1071 	abort = 0;
1072 abort:
1073 	err = xenbus_transaction_end(xbt, abort);
1074 	if (err) {
1075 		if (err == -EAGAIN && !abort)
1076 			goto again;
1077 		pr_warn("%s cannot complete xenstore transaction\n", __func__);
1078 		return err;
1079 	}
1080 
1081 	if (abort)
1082 		return -EFAULT;
1083 
1084 	xenbus_switch_state(dev, XenbusStateInitWait);
1085 
1086 	return 0;
1087 }
1088 
1089 static void set_backend_state(struct xenbus_device *dev,
1090 			      enum xenbus_state state)
1091 {
1092 	while (dev->state != state) {
1093 		switch (dev->state) {
1094 		case XenbusStateClosed:
1095 			switch (state) {
1096 			case XenbusStateInitWait:
1097 			case XenbusStateConnected:
1098 				xenbus_switch_state(dev, XenbusStateInitWait);
1099 				break;
1100 			case XenbusStateClosing:
1101 				xenbus_switch_state(dev, XenbusStateClosing);
1102 				break;
1103 			default:
1104 				WARN_ON(1);
1105 			}
1106 			break;
1107 		case XenbusStateInitWait:
1108 		case XenbusStateInitialised:
1109 			switch (state) {
1110 			case XenbusStateConnected:
1111 				if (backend_connect(dev))
1112 					return;
1113 				xenbus_switch_state(dev, XenbusStateConnected);
1114 				break;
1115 			case XenbusStateClosing:
1116 			case XenbusStateClosed:
1117 				xenbus_switch_state(dev, XenbusStateClosing);
1118 				break;
1119 			default:
1120 				WARN_ON(1);
1121 			}
1122 			break;
1123 		case XenbusStateConnected:
1124 			switch (state) {
1125 			case XenbusStateInitWait:
1126 			case XenbusStateClosing:
1127 			case XenbusStateClosed:
1128 				down(&pvcalls_back_global.frontends_lock);
1129 				backend_disconnect(dev);
1130 				up(&pvcalls_back_global.frontends_lock);
1131 				xenbus_switch_state(dev, XenbusStateClosing);
1132 				break;
1133 			default:
1134 				WARN_ON(1);
1135 			}
1136 			break;
1137 		case XenbusStateClosing:
1138 			switch (state) {
1139 			case XenbusStateInitWait:
1140 			case XenbusStateConnected:
1141 			case XenbusStateClosed:
1142 				xenbus_switch_state(dev, XenbusStateClosed);
1143 				break;
1144 			default:
1145 				WARN_ON(1);
1146 			}
1147 			break;
1148 		default:
1149 			WARN_ON(1);
1150 		}
1151 	}
1152 }
1153 
1154 static void pvcalls_back_changed(struct xenbus_device *dev,
1155 				 enum xenbus_state frontend_state)
1156 {
1157 	switch (frontend_state) {
1158 	case XenbusStateInitialising:
1159 		set_backend_state(dev, XenbusStateInitWait);
1160 		break;
1161 
1162 	case XenbusStateInitialised:
1163 	case XenbusStateConnected:
1164 		set_backend_state(dev, XenbusStateConnected);
1165 		break;
1166 
1167 	case XenbusStateClosing:
1168 		set_backend_state(dev, XenbusStateClosing);
1169 		break;
1170 
1171 	case XenbusStateClosed:
1172 		set_backend_state(dev, XenbusStateClosed);
1173 		if (xenbus_dev_is_online(dev))
1174 			break;
1175 		device_unregister(&dev->dev);
1176 		break;
1177 	case XenbusStateUnknown:
1178 		set_backend_state(dev, XenbusStateClosed);
1179 		device_unregister(&dev->dev);
1180 		break;
1181 
1182 	default:
1183 		xenbus_dev_fatal(dev, -EINVAL, "saw state %d at frontend",
1184 				 frontend_state);
1185 		break;
1186 	}
1187 }
1188 
1189 static void pvcalls_back_remove(struct xenbus_device *dev)
1190 {
1191 }
1192 
1193 static int pvcalls_back_uevent(const struct xenbus_device *xdev,
1194 			       struct kobj_uevent_env *env)
1195 {
1196 	return 0;
1197 }
1198 
1199 static const struct xenbus_device_id pvcalls_back_ids[] = {
1200 	{ "pvcalls" },
1201 	{ "" }
1202 };
1203 
1204 static struct xenbus_driver pvcalls_back_driver = {
1205 	.ids = pvcalls_back_ids,
1206 	.probe = pvcalls_back_probe,
1207 	.remove = pvcalls_back_remove,
1208 	.uevent = pvcalls_back_uevent,
1209 	.otherend_changed = pvcalls_back_changed,
1210 };
1211 
1212 static int __init pvcalls_back_init(void)
1213 {
1214 	int ret;
1215 
1216 	if (!xen_domain())
1217 		return -ENODEV;
1218 
1219 	ret = xenbus_register_backend(&pvcalls_back_driver);
1220 	if (ret < 0)
1221 		return ret;
1222 
1223 	sema_init(&pvcalls_back_global.frontends_lock, 1);
1224 	INIT_LIST_HEAD(&pvcalls_back_global.frontends);
1225 	return 0;
1226 }
1227 module_init(pvcalls_back_init);
1228 
1229 static void __exit pvcalls_back_fin(void)
1230 {
1231 	struct pvcalls_fedata *fedata, *nfedata;
1232 
1233 	down(&pvcalls_back_global.frontends_lock);
1234 	list_for_each_entry_safe(fedata, nfedata,
1235 				 &pvcalls_back_global.frontends, list) {
1236 		backend_disconnect(fedata->dev);
1237 	}
1238 	up(&pvcalls_back_global.frontends_lock);
1239 
1240 	xenbus_unregister_driver(&pvcalls_back_driver);
1241 }
1242 
1243 module_exit(pvcalls_back_fin);
1244 
1245 MODULE_DESCRIPTION("Xen PV Calls backend driver");
1246 MODULE_AUTHOR("Stefano Stabellini <sstabellini@kernel.org>");
1247 MODULE_LICENSE("GPL");
1248