xref: /openbmc/linux/drivers/xen/pvcalls-back.c (revision 2f61c664)
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 
18 #include <xen/events.h>
19 #include <xen/grant_table.h>
20 #include <xen/xen.h>
21 #include <xen/xenbus.h>
22 #include <xen/interface/io/pvcalls.h>
23 
24 #define PVCALLS_VERSIONS "1"
25 #define MAX_RING_ORDER XENBUS_MAX_RING_GRANT_ORDER
26 
27 static struct pvcalls_back_global {
28 	struct list_head frontends;
29 	struct semaphore frontends_lock;
30 } pvcalls_back_global;
31 
32 /*
33  * Per-frontend data structure. It contains pointers to the command
34  * ring, its event channel, a list of active sockets and a tree of
35  * passive sockets.
36  */
37 struct pvcalls_fedata {
38 	struct list_head list;
39 	struct xenbus_device *dev;
40 	struct xen_pvcalls_sring *sring;
41 	struct xen_pvcalls_back_ring ring;
42 	int irq;
43 	struct list_head socket_mappings;
44 	struct radix_tree_root socketpass_mappings;
45 	struct semaphore socket_lock;
46 };
47 
48 struct pvcalls_ioworker {
49 	struct work_struct register_work;
50 	struct workqueue_struct *wq;
51 };
52 
53 struct sock_mapping {
54 	struct list_head list;
55 	struct pvcalls_fedata *fedata;
56 	struct sockpass_mapping *sockpass;
57 	struct socket *sock;
58 	uint64_t id;
59 	grant_ref_t ref;
60 	struct pvcalls_data_intf *ring;
61 	void *bytes;
62 	struct pvcalls_data data;
63 	uint32_t ring_order;
64 	int irq;
65 	atomic_t read;
66 	atomic_t write;
67 	atomic_t io;
68 	atomic_t release;
69 	atomic_t eoi;
70 	void (*saved_data_ready)(struct sock *sk);
71 	struct pvcalls_ioworker ioworker;
72 };
73 
74 struct sockpass_mapping {
75 	struct list_head list;
76 	struct pvcalls_fedata *fedata;
77 	struct socket *sock;
78 	uint64_t id;
79 	struct xen_pvcalls_request reqcopy;
80 	spinlock_t copy_lock;
81 	struct workqueue_struct *wq;
82 	struct work_struct register_work;
83 	void (*saved_data_ready)(struct sock *sk);
84 };
85 
86 static irqreturn_t pvcalls_back_conn_event(int irq, void *sock_map);
87 static int pvcalls_back_release_active(struct xenbus_device *dev,
88 				       struct pvcalls_fedata *fedata,
89 				       struct sock_mapping *map);
90 
91 static bool pvcalls_conn_back_read(void *opaque)
92 {
93 	struct sock_mapping *map = (struct sock_mapping *)opaque;
94 	struct msghdr msg;
95 	struct kvec vec[2];
96 	RING_IDX cons, prod, size, wanted, array_size, masked_prod, masked_cons;
97 	int32_t error;
98 	struct pvcalls_data_intf *intf = map->ring;
99 	struct pvcalls_data *data = &map->data;
100 	unsigned long flags;
101 	int ret;
102 
103 	array_size = XEN_FLEX_RING_SIZE(map->ring_order);
104 	cons = intf->in_cons;
105 	prod = intf->in_prod;
106 	error = intf->in_error;
107 	/* read the indexes first, then deal with the data */
108 	virt_mb();
109 
110 	if (error)
111 		return false;
112 
113 	size = pvcalls_queued(prod, cons, array_size);
114 	if (size >= array_size)
115 		return false;
116 	spin_lock_irqsave(&map->sock->sk->sk_receive_queue.lock, flags);
117 	if (skb_queue_empty(&map->sock->sk->sk_receive_queue)) {
118 		atomic_set(&map->read, 0);
119 		spin_unlock_irqrestore(&map->sock->sk->sk_receive_queue.lock,
120 				flags);
121 		return true;
122 	}
123 	spin_unlock_irqrestore(&map->sock->sk->sk_receive_queue.lock, flags);
124 	wanted = array_size - size;
125 	masked_prod = pvcalls_mask(prod, array_size);
126 	masked_cons = pvcalls_mask(cons, array_size);
127 
128 	memset(&msg, 0, sizeof(msg));
129 	if (masked_prod < masked_cons) {
130 		vec[0].iov_base = data->in + masked_prod;
131 		vec[0].iov_len = wanted;
132 		iov_iter_kvec(&msg.msg_iter, WRITE, vec, 1, wanted);
133 	} else {
134 		vec[0].iov_base = data->in + masked_prod;
135 		vec[0].iov_len = array_size - masked_prod;
136 		vec[1].iov_base = data->in;
137 		vec[1].iov_len = wanted - vec[0].iov_len;
138 		iov_iter_kvec(&msg.msg_iter, WRITE, vec, 2, wanted);
139 	}
140 
141 	atomic_set(&map->read, 0);
142 	ret = inet_recvmsg(map->sock, &msg, wanted, MSG_DONTWAIT);
143 	WARN_ON(ret > wanted);
144 	if (ret == -EAGAIN) /* shouldn't happen */
145 		return true;
146 	if (!ret)
147 		ret = -ENOTCONN;
148 	spin_lock_irqsave(&map->sock->sk->sk_receive_queue.lock, flags);
149 	if (ret > 0 && !skb_queue_empty(&map->sock->sk->sk_receive_queue))
150 		atomic_inc(&map->read);
151 	spin_unlock_irqrestore(&map->sock->sk->sk_receive_queue.lock, flags);
152 
153 	/* write the data, then modify the indexes */
154 	virt_wmb();
155 	if (ret < 0) {
156 		atomic_set(&map->read, 0);
157 		intf->in_error = ret;
158 	} else
159 		intf->in_prod = prod + ret;
160 	/* update the indexes, then notify the other end */
161 	virt_wmb();
162 	notify_remote_via_irq(map->irq);
163 
164 	return true;
165 }
166 
167 static bool pvcalls_conn_back_write(struct sock_mapping *map)
168 {
169 	struct pvcalls_data_intf *intf = map->ring;
170 	struct pvcalls_data *data = &map->data;
171 	struct msghdr msg;
172 	struct kvec vec[2];
173 	RING_IDX cons, prod, size, array_size;
174 	int ret;
175 
176 	cons = intf->out_cons;
177 	prod = intf->out_prod;
178 	/* read the indexes before dealing with the data */
179 	virt_mb();
180 
181 	array_size = XEN_FLEX_RING_SIZE(map->ring_order);
182 	size = pvcalls_queued(prod, cons, array_size);
183 	if (size == 0)
184 		return false;
185 
186 	memset(&msg, 0, sizeof(msg));
187 	msg.msg_flags |= MSG_DONTWAIT;
188 	if (pvcalls_mask(prod, array_size) > pvcalls_mask(cons, array_size)) {
189 		vec[0].iov_base = data->out + pvcalls_mask(cons, array_size);
190 		vec[0].iov_len = size;
191 		iov_iter_kvec(&msg.msg_iter, READ, vec, 1, size);
192 	} else {
193 		vec[0].iov_base = data->out + pvcalls_mask(cons, array_size);
194 		vec[0].iov_len = array_size - pvcalls_mask(cons, array_size);
195 		vec[1].iov_base = data->out;
196 		vec[1].iov_len = size - vec[0].iov_len;
197 		iov_iter_kvec(&msg.msg_iter, READ, vec, 2, size);
198 	}
199 
200 	atomic_set(&map->write, 0);
201 	ret = inet_sendmsg(map->sock, &msg, size);
202 	if (ret == -EAGAIN) {
203 		atomic_inc(&map->write);
204 		atomic_inc(&map->io);
205 		return true;
206 	}
207 
208 	/* write the data, then update the indexes */
209 	virt_wmb();
210 	if (ret < 0) {
211 		intf->out_error = ret;
212 	} else {
213 		intf->out_error = 0;
214 		intf->out_cons = cons + ret;
215 		prod = intf->out_prod;
216 	}
217 	/* update the indexes, then notify the other end */
218 	virt_wmb();
219 	if (prod != cons + ret) {
220 		atomic_inc(&map->write);
221 		atomic_inc(&map->io);
222 	}
223 	notify_remote_via_irq(map->irq);
224 
225 	return true;
226 }
227 
228 static void pvcalls_back_ioworker(struct work_struct *work)
229 {
230 	struct pvcalls_ioworker *ioworker = container_of(work,
231 		struct pvcalls_ioworker, register_work);
232 	struct sock_mapping *map = container_of(ioworker, struct sock_mapping,
233 		ioworker);
234 	unsigned int eoi_flags = XEN_EOI_FLAG_SPURIOUS;
235 
236 	while (atomic_read(&map->io) > 0) {
237 		if (atomic_read(&map->release) > 0) {
238 			atomic_set(&map->release, 0);
239 			return;
240 		}
241 
242 		if (atomic_read(&map->read) > 0 &&
243 		    pvcalls_conn_back_read(map))
244 			eoi_flags = 0;
245 		if (atomic_read(&map->write) > 0 &&
246 		    pvcalls_conn_back_write(map))
247 			eoi_flags = 0;
248 
249 		if (atomic_read(&map->eoi) > 0 && !atomic_read(&map->write)) {
250 			atomic_set(&map->eoi, 0);
251 			xen_irq_lateeoi(map->irq, eoi_flags);
252 			eoi_flags = XEN_EOI_FLAG_SPURIOUS;
253 		}
254 
255 		atomic_dec(&map->io);
256 	}
257 }
258 
259 static int pvcalls_back_socket(struct xenbus_device *dev,
260 		struct xen_pvcalls_request *req)
261 {
262 	struct pvcalls_fedata *fedata;
263 	int ret;
264 	struct xen_pvcalls_response *rsp;
265 
266 	fedata = dev_get_drvdata(&dev->dev);
267 
268 	if (req->u.socket.domain != AF_INET ||
269 	    req->u.socket.type != SOCK_STREAM ||
270 	    (req->u.socket.protocol != IPPROTO_IP &&
271 	     req->u.socket.protocol != AF_INET))
272 		ret = -EAFNOSUPPORT;
273 	else
274 		ret = 0;
275 
276 	/* leave the actual socket allocation for later */
277 
278 	rsp = RING_GET_RESPONSE(&fedata->ring, fedata->ring.rsp_prod_pvt++);
279 	rsp->req_id = req->req_id;
280 	rsp->cmd = req->cmd;
281 	rsp->u.socket.id = req->u.socket.id;
282 	rsp->ret = ret;
283 
284 	return 0;
285 }
286 
287 static void pvcalls_sk_state_change(struct sock *sock)
288 {
289 	struct sock_mapping *map = sock->sk_user_data;
290 
291 	if (map == NULL)
292 		return;
293 
294 	atomic_inc(&map->read);
295 	notify_remote_via_irq(map->irq);
296 }
297 
298 static void pvcalls_sk_data_ready(struct sock *sock)
299 {
300 	struct sock_mapping *map = sock->sk_user_data;
301 	struct pvcalls_ioworker *iow;
302 
303 	if (map == NULL)
304 		return;
305 
306 	iow = &map->ioworker;
307 	atomic_inc(&map->read);
308 	atomic_inc(&map->io);
309 	queue_work(iow->wq, &iow->register_work);
310 }
311 
312 static struct sock_mapping *pvcalls_new_active_socket(
313 		struct pvcalls_fedata *fedata,
314 		uint64_t id,
315 		grant_ref_t ref,
316 		evtchn_port_t evtchn,
317 		struct socket *sock)
318 {
319 	int ret;
320 	struct sock_mapping *map;
321 	void *page;
322 
323 	map = kzalloc(sizeof(*map), GFP_KERNEL);
324 	if (map == NULL)
325 		return NULL;
326 
327 	map->fedata = fedata;
328 	map->sock = sock;
329 	map->id = id;
330 	map->ref = ref;
331 
332 	ret = xenbus_map_ring_valloc(fedata->dev, &ref, 1, &page);
333 	if (ret < 0)
334 		goto out;
335 	map->ring = page;
336 	map->ring_order = map->ring->ring_order;
337 	/* first read the order, then map the data ring */
338 	virt_rmb();
339 	if (map->ring_order > MAX_RING_ORDER) {
340 		pr_warn("%s frontend requested ring_order %u, which is > MAX (%u)\n",
341 				__func__, map->ring_order, MAX_RING_ORDER);
342 		goto out;
343 	}
344 	ret = xenbus_map_ring_valloc(fedata->dev, map->ring->ref,
345 				     (1 << map->ring_order), &page);
346 	if (ret < 0)
347 		goto out;
348 	map->bytes = page;
349 
350 	ret = bind_interdomain_evtchn_to_irqhandler_lateeoi(
351 			fedata->dev, evtchn,
352 			pvcalls_back_conn_event, 0, "pvcalls-backend", map);
353 	if (ret < 0)
354 		goto out;
355 	map->irq = ret;
356 
357 	map->data.in = map->bytes;
358 	map->data.out = map->bytes + XEN_FLEX_RING_SIZE(map->ring_order);
359 
360 	map->ioworker.wq = alloc_workqueue("pvcalls_io", WQ_UNBOUND, 1);
361 	if (!map->ioworker.wq)
362 		goto out;
363 	atomic_set(&map->io, 1);
364 	INIT_WORK(&map->ioworker.register_work,	pvcalls_back_ioworker);
365 
366 	down(&fedata->socket_lock);
367 	list_add_tail(&map->list, &fedata->socket_mappings);
368 	up(&fedata->socket_lock);
369 
370 	write_lock_bh(&map->sock->sk->sk_callback_lock);
371 	map->saved_data_ready = map->sock->sk->sk_data_ready;
372 	map->sock->sk->sk_user_data = map;
373 	map->sock->sk->sk_data_ready = pvcalls_sk_data_ready;
374 	map->sock->sk->sk_state_change = pvcalls_sk_state_change;
375 	write_unlock_bh(&map->sock->sk->sk_callback_lock);
376 
377 	return map;
378 out:
379 	down(&fedata->socket_lock);
380 	list_del(&map->list);
381 	pvcalls_back_release_active(fedata->dev, fedata, map);
382 	up(&fedata->socket_lock);
383 	return NULL;
384 }
385 
386 static int pvcalls_back_connect(struct xenbus_device *dev,
387 				struct xen_pvcalls_request *req)
388 {
389 	struct pvcalls_fedata *fedata;
390 	int ret = -EINVAL;
391 	struct socket *sock;
392 	struct sock_mapping *map;
393 	struct xen_pvcalls_response *rsp;
394 	struct sockaddr *sa = (struct sockaddr *)&req->u.connect.addr;
395 
396 	fedata = dev_get_drvdata(&dev->dev);
397 
398 	if (req->u.connect.len < sizeof(sa->sa_family) ||
399 	    req->u.connect.len > sizeof(req->u.connect.addr) ||
400 	    sa->sa_family != AF_INET)
401 		goto out;
402 
403 	ret = sock_create(AF_INET, SOCK_STREAM, 0, &sock);
404 	if (ret < 0)
405 		goto out;
406 	ret = inet_stream_connect(sock, sa, req->u.connect.len, 0);
407 	if (ret < 0) {
408 		sock_release(sock);
409 		goto out;
410 	}
411 
412 	map = pvcalls_new_active_socket(fedata,
413 					req->u.connect.id,
414 					req->u.connect.ref,
415 					req->u.connect.evtchn,
416 					sock);
417 	if (!map) {
418 		ret = -EFAULT;
419 		sock_release(sock);
420 	}
421 
422 out:
423 	rsp = RING_GET_RESPONSE(&fedata->ring, fedata->ring.rsp_prod_pvt++);
424 	rsp->req_id = req->req_id;
425 	rsp->cmd = req->cmd;
426 	rsp->u.connect.id = req->u.connect.id;
427 	rsp->ret = ret;
428 
429 	return 0;
430 }
431 
432 static int pvcalls_back_release_active(struct xenbus_device *dev,
433 				       struct pvcalls_fedata *fedata,
434 				       struct sock_mapping *map)
435 {
436 	disable_irq(map->irq);
437 	if (map->sock->sk != NULL) {
438 		write_lock_bh(&map->sock->sk->sk_callback_lock);
439 		map->sock->sk->sk_user_data = NULL;
440 		map->sock->sk->sk_data_ready = map->saved_data_ready;
441 		write_unlock_bh(&map->sock->sk->sk_callback_lock);
442 	}
443 
444 	atomic_set(&map->release, 1);
445 	flush_work(&map->ioworker.register_work);
446 
447 	xenbus_unmap_ring_vfree(dev, map->bytes);
448 	xenbus_unmap_ring_vfree(dev, (void *)map->ring);
449 	unbind_from_irqhandler(map->irq, map);
450 
451 	sock_release(map->sock);
452 	kfree(map);
453 
454 	return 0;
455 }
456 
457 static int pvcalls_back_release_passive(struct xenbus_device *dev,
458 					struct pvcalls_fedata *fedata,
459 					struct sockpass_mapping *mappass)
460 {
461 	if (mappass->sock->sk != NULL) {
462 		write_lock_bh(&mappass->sock->sk->sk_callback_lock);
463 		mappass->sock->sk->sk_user_data = NULL;
464 		mappass->sock->sk->sk_data_ready = mappass->saved_data_ready;
465 		write_unlock_bh(&mappass->sock->sk->sk_callback_lock);
466 	}
467 	sock_release(mappass->sock);
468 	flush_workqueue(mappass->wq);
469 	destroy_workqueue(mappass->wq);
470 	kfree(mappass);
471 
472 	return 0;
473 }
474 
475 static int pvcalls_back_release(struct xenbus_device *dev,
476 				struct xen_pvcalls_request *req)
477 {
478 	struct pvcalls_fedata *fedata;
479 	struct sock_mapping *map, *n;
480 	struct sockpass_mapping *mappass;
481 	int ret = 0;
482 	struct xen_pvcalls_response *rsp;
483 
484 	fedata = dev_get_drvdata(&dev->dev);
485 
486 	down(&fedata->socket_lock);
487 	list_for_each_entry_safe(map, n, &fedata->socket_mappings, list) {
488 		if (map->id == req->u.release.id) {
489 			list_del(&map->list);
490 			up(&fedata->socket_lock);
491 			ret = pvcalls_back_release_active(dev, fedata, map);
492 			goto out;
493 		}
494 	}
495 	mappass = radix_tree_lookup(&fedata->socketpass_mappings,
496 				    req->u.release.id);
497 	if (mappass != NULL) {
498 		radix_tree_delete(&fedata->socketpass_mappings, mappass->id);
499 		up(&fedata->socket_lock);
500 		ret = pvcalls_back_release_passive(dev, fedata, mappass);
501 	} else
502 		up(&fedata->socket_lock);
503 
504 out:
505 	rsp = RING_GET_RESPONSE(&fedata->ring, fedata->ring.rsp_prod_pvt++);
506 	rsp->req_id = req->req_id;
507 	rsp->u.release.id = req->u.release.id;
508 	rsp->cmd = req->cmd;
509 	rsp->ret = ret;
510 	return 0;
511 }
512 
513 static void __pvcalls_back_accept(struct work_struct *work)
514 {
515 	struct sockpass_mapping *mappass = container_of(
516 		work, struct sockpass_mapping, register_work);
517 	struct sock_mapping *map;
518 	struct pvcalls_ioworker *iow;
519 	struct pvcalls_fedata *fedata;
520 	struct socket *sock;
521 	struct xen_pvcalls_response *rsp;
522 	struct xen_pvcalls_request *req;
523 	int notify;
524 	int ret = -EINVAL;
525 	unsigned long flags;
526 
527 	fedata = mappass->fedata;
528 	/*
529 	 * __pvcalls_back_accept can race against pvcalls_back_accept.
530 	 * We only need to check the value of "cmd" on read. It could be
531 	 * done atomically, but to simplify the code on the write side, we
532 	 * use a spinlock.
533 	 */
534 	spin_lock_irqsave(&mappass->copy_lock, flags);
535 	req = &mappass->reqcopy;
536 	if (req->cmd != PVCALLS_ACCEPT) {
537 		spin_unlock_irqrestore(&mappass->copy_lock, flags);
538 		return;
539 	}
540 	spin_unlock_irqrestore(&mappass->copy_lock, flags);
541 
542 	sock = sock_alloc();
543 	if (sock == NULL)
544 		goto out_error;
545 	sock->type = mappass->sock->type;
546 	sock->ops = mappass->sock->ops;
547 
548 	ret = inet_accept(mappass->sock, sock, O_NONBLOCK, true);
549 	if (ret == -EAGAIN) {
550 		sock_release(sock);
551 		return;
552 	}
553 
554 	map = pvcalls_new_active_socket(fedata,
555 					req->u.accept.id_new,
556 					req->u.accept.ref,
557 					req->u.accept.evtchn,
558 					sock);
559 	if (!map) {
560 		ret = -EFAULT;
561 		sock_release(sock);
562 		goto out_error;
563 	}
564 
565 	map->sockpass = mappass;
566 	iow = &map->ioworker;
567 	atomic_inc(&map->read);
568 	atomic_inc(&map->io);
569 	queue_work(iow->wq, &iow->register_work);
570 
571 out_error:
572 	rsp = RING_GET_RESPONSE(&fedata->ring, fedata->ring.rsp_prod_pvt++);
573 	rsp->req_id = req->req_id;
574 	rsp->cmd = req->cmd;
575 	rsp->u.accept.id = req->u.accept.id;
576 	rsp->ret = ret;
577 	RING_PUSH_RESPONSES_AND_CHECK_NOTIFY(&fedata->ring, notify);
578 	if (notify)
579 		notify_remote_via_irq(fedata->irq);
580 
581 	mappass->reqcopy.cmd = 0;
582 }
583 
584 static void pvcalls_pass_sk_data_ready(struct sock *sock)
585 {
586 	struct sockpass_mapping *mappass = sock->sk_user_data;
587 	struct pvcalls_fedata *fedata;
588 	struct xen_pvcalls_response *rsp;
589 	unsigned long flags;
590 	int notify;
591 
592 	if (mappass == NULL)
593 		return;
594 
595 	fedata = mappass->fedata;
596 	spin_lock_irqsave(&mappass->copy_lock, flags);
597 	if (mappass->reqcopy.cmd == PVCALLS_POLL) {
598 		rsp = RING_GET_RESPONSE(&fedata->ring,
599 					fedata->ring.rsp_prod_pvt++);
600 		rsp->req_id = mappass->reqcopy.req_id;
601 		rsp->u.poll.id = mappass->reqcopy.u.poll.id;
602 		rsp->cmd = mappass->reqcopy.cmd;
603 		rsp->ret = 0;
604 
605 		mappass->reqcopy.cmd = 0;
606 		spin_unlock_irqrestore(&mappass->copy_lock, flags);
607 
608 		RING_PUSH_RESPONSES_AND_CHECK_NOTIFY(&fedata->ring, notify);
609 		if (notify)
610 			notify_remote_via_irq(mappass->fedata->irq);
611 	} else {
612 		spin_unlock_irqrestore(&mappass->copy_lock, flags);
613 		queue_work(mappass->wq, &mappass->register_work);
614 	}
615 }
616 
617 static int pvcalls_back_bind(struct xenbus_device *dev,
618 			     struct xen_pvcalls_request *req)
619 {
620 	struct pvcalls_fedata *fedata;
621 	int ret;
622 	struct sockpass_mapping *map;
623 	struct xen_pvcalls_response *rsp;
624 
625 	fedata = dev_get_drvdata(&dev->dev);
626 
627 	map = kzalloc(sizeof(*map), GFP_KERNEL);
628 	if (map == NULL) {
629 		ret = -ENOMEM;
630 		goto out;
631 	}
632 
633 	INIT_WORK(&map->register_work, __pvcalls_back_accept);
634 	spin_lock_init(&map->copy_lock);
635 	map->wq = alloc_workqueue("pvcalls_wq", WQ_UNBOUND, 1);
636 	if (!map->wq) {
637 		ret = -ENOMEM;
638 		goto out;
639 	}
640 
641 	ret = sock_create(AF_INET, SOCK_STREAM, 0, &map->sock);
642 	if (ret < 0)
643 		goto out;
644 
645 	ret = inet_bind(map->sock, (struct sockaddr *)&req->u.bind.addr,
646 			req->u.bind.len);
647 	if (ret < 0)
648 		goto out;
649 
650 	map->fedata = fedata;
651 	map->id = req->u.bind.id;
652 
653 	down(&fedata->socket_lock);
654 	ret = radix_tree_insert(&fedata->socketpass_mappings, map->id,
655 				map);
656 	up(&fedata->socket_lock);
657 	if (ret)
658 		goto out;
659 
660 	write_lock_bh(&map->sock->sk->sk_callback_lock);
661 	map->saved_data_ready = map->sock->sk->sk_data_ready;
662 	map->sock->sk->sk_user_data = map;
663 	map->sock->sk->sk_data_ready = pvcalls_pass_sk_data_ready;
664 	write_unlock_bh(&map->sock->sk->sk_callback_lock);
665 
666 out:
667 	if (ret) {
668 		if (map && map->sock)
669 			sock_release(map->sock);
670 		if (map && map->wq)
671 			destroy_workqueue(map->wq);
672 		kfree(map);
673 	}
674 	rsp = RING_GET_RESPONSE(&fedata->ring, fedata->ring.rsp_prod_pvt++);
675 	rsp->req_id = req->req_id;
676 	rsp->cmd = req->cmd;
677 	rsp->u.bind.id = req->u.bind.id;
678 	rsp->ret = ret;
679 	return 0;
680 }
681 
682 static int pvcalls_back_listen(struct xenbus_device *dev,
683 			       struct xen_pvcalls_request *req)
684 {
685 	struct pvcalls_fedata *fedata;
686 	int ret = -EINVAL;
687 	struct sockpass_mapping *map;
688 	struct xen_pvcalls_response *rsp;
689 
690 	fedata = dev_get_drvdata(&dev->dev);
691 
692 	down(&fedata->socket_lock);
693 	map = radix_tree_lookup(&fedata->socketpass_mappings, req->u.listen.id);
694 	up(&fedata->socket_lock);
695 	if (map == NULL)
696 		goto out;
697 
698 	ret = inet_listen(map->sock, req->u.listen.backlog);
699 
700 out:
701 	rsp = RING_GET_RESPONSE(&fedata->ring, fedata->ring.rsp_prod_pvt++);
702 	rsp->req_id = req->req_id;
703 	rsp->cmd = req->cmd;
704 	rsp->u.listen.id = req->u.listen.id;
705 	rsp->ret = ret;
706 	return 0;
707 }
708 
709 static int pvcalls_back_accept(struct xenbus_device *dev,
710 			       struct xen_pvcalls_request *req)
711 {
712 	struct pvcalls_fedata *fedata;
713 	struct sockpass_mapping *mappass;
714 	int ret = -EINVAL;
715 	struct xen_pvcalls_response *rsp;
716 	unsigned long flags;
717 
718 	fedata = dev_get_drvdata(&dev->dev);
719 
720 	down(&fedata->socket_lock);
721 	mappass = radix_tree_lookup(&fedata->socketpass_mappings,
722 		req->u.accept.id);
723 	up(&fedata->socket_lock);
724 	if (mappass == NULL)
725 		goto out_error;
726 
727 	/*
728 	 * Limitation of the current implementation: only support one
729 	 * concurrent accept or poll call on one socket.
730 	 */
731 	spin_lock_irqsave(&mappass->copy_lock, flags);
732 	if (mappass->reqcopy.cmd != 0) {
733 		spin_unlock_irqrestore(&mappass->copy_lock, flags);
734 		ret = -EINTR;
735 		goto out_error;
736 	}
737 
738 	mappass->reqcopy = *req;
739 	spin_unlock_irqrestore(&mappass->copy_lock, flags);
740 	queue_work(mappass->wq, &mappass->register_work);
741 
742 	/* Tell the caller we don't need to send back a notification yet */
743 	return -1;
744 
745 out_error:
746 	rsp = RING_GET_RESPONSE(&fedata->ring, fedata->ring.rsp_prod_pvt++);
747 	rsp->req_id = req->req_id;
748 	rsp->cmd = req->cmd;
749 	rsp->u.accept.id = req->u.accept.id;
750 	rsp->ret = ret;
751 	return 0;
752 }
753 
754 static int pvcalls_back_poll(struct xenbus_device *dev,
755 			     struct xen_pvcalls_request *req)
756 {
757 	struct pvcalls_fedata *fedata;
758 	struct sockpass_mapping *mappass;
759 	struct xen_pvcalls_response *rsp;
760 	struct inet_connection_sock *icsk;
761 	struct request_sock_queue *queue;
762 	unsigned long flags;
763 	int ret;
764 	bool data;
765 
766 	fedata = dev_get_drvdata(&dev->dev);
767 
768 	down(&fedata->socket_lock);
769 	mappass = radix_tree_lookup(&fedata->socketpass_mappings,
770 				    req->u.poll.id);
771 	up(&fedata->socket_lock);
772 	if (mappass == NULL)
773 		return -EINVAL;
774 
775 	/*
776 	 * Limitation of the current implementation: only support one
777 	 * concurrent accept or poll call on one socket.
778 	 */
779 	spin_lock_irqsave(&mappass->copy_lock, flags);
780 	if (mappass->reqcopy.cmd != 0) {
781 		ret = -EINTR;
782 		goto out;
783 	}
784 
785 	mappass->reqcopy = *req;
786 	icsk = inet_csk(mappass->sock->sk);
787 	queue = &icsk->icsk_accept_queue;
788 	data = READ_ONCE(queue->rskq_accept_head) != NULL;
789 	if (data) {
790 		mappass->reqcopy.cmd = 0;
791 		ret = 0;
792 		goto out;
793 	}
794 	spin_unlock_irqrestore(&mappass->copy_lock, flags);
795 
796 	/* Tell the caller we don't need to send back a notification yet */
797 	return -1;
798 
799 out:
800 	spin_unlock_irqrestore(&mappass->copy_lock, flags);
801 
802 	rsp = RING_GET_RESPONSE(&fedata->ring, fedata->ring.rsp_prod_pvt++);
803 	rsp->req_id = req->req_id;
804 	rsp->cmd = req->cmd;
805 	rsp->u.poll.id = req->u.poll.id;
806 	rsp->ret = ret;
807 	return 0;
808 }
809 
810 static int pvcalls_back_handle_cmd(struct xenbus_device *dev,
811 				   struct xen_pvcalls_request *req)
812 {
813 	int ret = 0;
814 
815 	switch (req->cmd) {
816 	case PVCALLS_SOCKET:
817 		ret = pvcalls_back_socket(dev, req);
818 		break;
819 	case PVCALLS_CONNECT:
820 		ret = pvcalls_back_connect(dev, req);
821 		break;
822 	case PVCALLS_RELEASE:
823 		ret = pvcalls_back_release(dev, req);
824 		break;
825 	case PVCALLS_BIND:
826 		ret = pvcalls_back_bind(dev, req);
827 		break;
828 	case PVCALLS_LISTEN:
829 		ret = pvcalls_back_listen(dev, req);
830 		break;
831 	case PVCALLS_ACCEPT:
832 		ret = pvcalls_back_accept(dev, req);
833 		break;
834 	case PVCALLS_POLL:
835 		ret = pvcalls_back_poll(dev, req);
836 		break;
837 	default:
838 	{
839 		struct pvcalls_fedata *fedata;
840 		struct xen_pvcalls_response *rsp;
841 
842 		fedata = dev_get_drvdata(&dev->dev);
843 		rsp = RING_GET_RESPONSE(
844 				&fedata->ring, fedata->ring.rsp_prod_pvt++);
845 		rsp->req_id = req->req_id;
846 		rsp->cmd = req->cmd;
847 		rsp->ret = -ENOTSUPP;
848 		break;
849 	}
850 	}
851 	return ret;
852 }
853 
854 static void pvcalls_back_work(struct pvcalls_fedata *fedata)
855 {
856 	int notify, notify_all = 0, more = 1;
857 	struct xen_pvcalls_request req;
858 	struct xenbus_device *dev = fedata->dev;
859 
860 	while (more) {
861 		while (RING_HAS_UNCONSUMED_REQUESTS(&fedata->ring)) {
862 			RING_COPY_REQUEST(&fedata->ring,
863 					  fedata->ring.req_cons++,
864 					  &req);
865 
866 			if (!pvcalls_back_handle_cmd(dev, &req)) {
867 				RING_PUSH_RESPONSES_AND_CHECK_NOTIFY(
868 					&fedata->ring, notify);
869 				notify_all += notify;
870 			}
871 		}
872 
873 		if (notify_all) {
874 			notify_remote_via_irq(fedata->irq);
875 			notify_all = 0;
876 		}
877 
878 		RING_FINAL_CHECK_FOR_REQUESTS(&fedata->ring, more);
879 	}
880 }
881 
882 static irqreturn_t pvcalls_back_event(int irq, void *dev_id)
883 {
884 	struct xenbus_device *dev = dev_id;
885 	struct pvcalls_fedata *fedata = NULL;
886 	unsigned int eoi_flags = XEN_EOI_FLAG_SPURIOUS;
887 
888 	if (dev) {
889 		fedata = dev_get_drvdata(&dev->dev);
890 		if (fedata) {
891 			pvcalls_back_work(fedata);
892 			eoi_flags = 0;
893 		}
894 	}
895 
896 	xen_irq_lateeoi(irq, eoi_flags);
897 
898 	return IRQ_HANDLED;
899 }
900 
901 static irqreturn_t pvcalls_back_conn_event(int irq, void *sock_map)
902 {
903 	struct sock_mapping *map = sock_map;
904 	struct pvcalls_ioworker *iow;
905 
906 	if (map == NULL || map->sock == NULL || map->sock->sk == NULL ||
907 		map->sock->sk->sk_user_data != map) {
908 		xen_irq_lateeoi(irq, 0);
909 		return IRQ_HANDLED;
910 	}
911 
912 	iow = &map->ioworker;
913 
914 	atomic_inc(&map->write);
915 	atomic_inc(&map->eoi);
916 	atomic_inc(&map->io);
917 	queue_work(iow->wq, &iow->register_work);
918 
919 	return IRQ_HANDLED;
920 }
921 
922 static int backend_connect(struct xenbus_device *dev)
923 {
924 	int err;
925 	evtchn_port_t evtchn;
926 	grant_ref_t ring_ref;
927 	struct pvcalls_fedata *fedata = NULL;
928 
929 	fedata = kzalloc(sizeof(struct pvcalls_fedata), GFP_KERNEL);
930 	if (!fedata)
931 		return -ENOMEM;
932 
933 	fedata->irq = -1;
934 	err = xenbus_scanf(XBT_NIL, dev->otherend, "port", "%u",
935 			   &evtchn);
936 	if (err != 1) {
937 		err = -EINVAL;
938 		xenbus_dev_fatal(dev, err, "reading %s/event-channel",
939 				 dev->otherend);
940 		goto error;
941 	}
942 
943 	err = xenbus_scanf(XBT_NIL, dev->otherend, "ring-ref", "%u", &ring_ref);
944 	if (err != 1) {
945 		err = -EINVAL;
946 		xenbus_dev_fatal(dev, err, "reading %s/ring-ref",
947 				 dev->otherend);
948 		goto error;
949 	}
950 
951 	err = bind_interdomain_evtchn_to_irq_lateeoi(dev, evtchn);
952 	if (err < 0)
953 		goto error;
954 	fedata->irq = err;
955 
956 	err = request_threaded_irq(fedata->irq, NULL, pvcalls_back_event,
957 				   IRQF_ONESHOT, "pvcalls-back", dev);
958 	if (err < 0)
959 		goto error;
960 
961 	err = xenbus_map_ring_valloc(dev, &ring_ref, 1,
962 				     (void **)&fedata->sring);
963 	if (err < 0)
964 		goto error;
965 
966 	BACK_RING_INIT(&fedata->ring, fedata->sring, XEN_PAGE_SIZE * 1);
967 	fedata->dev = dev;
968 
969 	INIT_LIST_HEAD(&fedata->socket_mappings);
970 	INIT_RADIX_TREE(&fedata->socketpass_mappings, GFP_KERNEL);
971 	sema_init(&fedata->socket_lock, 1);
972 	dev_set_drvdata(&dev->dev, fedata);
973 
974 	down(&pvcalls_back_global.frontends_lock);
975 	list_add_tail(&fedata->list, &pvcalls_back_global.frontends);
976 	up(&pvcalls_back_global.frontends_lock);
977 
978 	return 0;
979 
980  error:
981 	if (fedata->irq >= 0)
982 		unbind_from_irqhandler(fedata->irq, dev);
983 	if (fedata->sring != NULL)
984 		xenbus_unmap_ring_vfree(dev, fedata->sring);
985 	kfree(fedata);
986 	return err;
987 }
988 
989 static int backend_disconnect(struct xenbus_device *dev)
990 {
991 	struct pvcalls_fedata *fedata;
992 	struct sock_mapping *map, *n;
993 	struct sockpass_mapping *mappass;
994 	struct radix_tree_iter iter;
995 	void **slot;
996 
997 
998 	fedata = dev_get_drvdata(&dev->dev);
999 
1000 	down(&fedata->socket_lock);
1001 	list_for_each_entry_safe(map, n, &fedata->socket_mappings, list) {
1002 		list_del(&map->list);
1003 		pvcalls_back_release_active(dev, fedata, map);
1004 	}
1005 
1006 	radix_tree_for_each_slot(slot, &fedata->socketpass_mappings, &iter, 0) {
1007 		mappass = radix_tree_deref_slot(slot);
1008 		if (!mappass)
1009 			continue;
1010 		if (radix_tree_exception(mappass)) {
1011 			if (radix_tree_deref_retry(mappass))
1012 				slot = radix_tree_iter_retry(&iter);
1013 		} else {
1014 			radix_tree_delete(&fedata->socketpass_mappings,
1015 					  mappass->id);
1016 			pvcalls_back_release_passive(dev, fedata, mappass);
1017 		}
1018 	}
1019 	up(&fedata->socket_lock);
1020 
1021 	unbind_from_irqhandler(fedata->irq, dev);
1022 	xenbus_unmap_ring_vfree(dev, fedata->sring);
1023 
1024 	list_del(&fedata->list);
1025 	kfree(fedata);
1026 	dev_set_drvdata(&dev->dev, NULL);
1027 
1028 	return 0;
1029 }
1030 
1031 static int pvcalls_back_probe(struct xenbus_device *dev,
1032 			      const struct xenbus_device_id *id)
1033 {
1034 	int err, abort;
1035 	struct xenbus_transaction xbt;
1036 
1037 again:
1038 	abort = 1;
1039 
1040 	err = xenbus_transaction_start(&xbt);
1041 	if (err) {
1042 		pr_warn("%s cannot create xenstore transaction\n", __func__);
1043 		return err;
1044 	}
1045 
1046 	err = xenbus_printf(xbt, dev->nodename, "versions", "%s",
1047 			    PVCALLS_VERSIONS);
1048 	if (err) {
1049 		pr_warn("%s write out 'versions' failed\n", __func__);
1050 		goto abort;
1051 	}
1052 
1053 	err = xenbus_printf(xbt, dev->nodename, "max-page-order", "%u",
1054 			    MAX_RING_ORDER);
1055 	if (err) {
1056 		pr_warn("%s write out 'max-page-order' failed\n", __func__);
1057 		goto abort;
1058 	}
1059 
1060 	err = xenbus_printf(xbt, dev->nodename, "function-calls",
1061 			    XENBUS_FUNCTIONS_CALLS);
1062 	if (err) {
1063 		pr_warn("%s write out 'function-calls' failed\n", __func__);
1064 		goto abort;
1065 	}
1066 
1067 	abort = 0;
1068 abort:
1069 	err = xenbus_transaction_end(xbt, abort);
1070 	if (err) {
1071 		if (err == -EAGAIN && !abort)
1072 			goto again;
1073 		pr_warn("%s cannot complete xenstore transaction\n", __func__);
1074 		return err;
1075 	}
1076 
1077 	if (abort)
1078 		return -EFAULT;
1079 
1080 	xenbus_switch_state(dev, XenbusStateInitWait);
1081 
1082 	return 0;
1083 }
1084 
1085 static void set_backend_state(struct xenbus_device *dev,
1086 			      enum xenbus_state state)
1087 {
1088 	while (dev->state != state) {
1089 		switch (dev->state) {
1090 		case XenbusStateClosed:
1091 			switch (state) {
1092 			case XenbusStateInitWait:
1093 			case XenbusStateConnected:
1094 				xenbus_switch_state(dev, XenbusStateInitWait);
1095 				break;
1096 			case XenbusStateClosing:
1097 				xenbus_switch_state(dev, XenbusStateClosing);
1098 				break;
1099 			default:
1100 				WARN_ON(1);
1101 			}
1102 			break;
1103 		case XenbusStateInitWait:
1104 		case XenbusStateInitialised:
1105 			switch (state) {
1106 			case XenbusStateConnected:
1107 				if (backend_connect(dev))
1108 					return;
1109 				xenbus_switch_state(dev, XenbusStateConnected);
1110 				break;
1111 			case XenbusStateClosing:
1112 			case XenbusStateClosed:
1113 				xenbus_switch_state(dev, XenbusStateClosing);
1114 				break;
1115 			default:
1116 				WARN_ON(1);
1117 			}
1118 			break;
1119 		case XenbusStateConnected:
1120 			switch (state) {
1121 			case XenbusStateInitWait:
1122 			case XenbusStateClosing:
1123 			case XenbusStateClosed:
1124 				down(&pvcalls_back_global.frontends_lock);
1125 				backend_disconnect(dev);
1126 				up(&pvcalls_back_global.frontends_lock);
1127 				xenbus_switch_state(dev, XenbusStateClosing);
1128 				break;
1129 			default:
1130 				WARN_ON(1);
1131 			}
1132 			break;
1133 		case XenbusStateClosing:
1134 			switch (state) {
1135 			case XenbusStateInitWait:
1136 			case XenbusStateConnected:
1137 			case XenbusStateClosed:
1138 				xenbus_switch_state(dev, XenbusStateClosed);
1139 				break;
1140 			default:
1141 				WARN_ON(1);
1142 			}
1143 			break;
1144 		default:
1145 			WARN_ON(1);
1146 		}
1147 	}
1148 }
1149 
1150 static void pvcalls_back_changed(struct xenbus_device *dev,
1151 				 enum xenbus_state frontend_state)
1152 {
1153 	switch (frontend_state) {
1154 	case XenbusStateInitialising:
1155 		set_backend_state(dev, XenbusStateInitWait);
1156 		break;
1157 
1158 	case XenbusStateInitialised:
1159 	case XenbusStateConnected:
1160 		set_backend_state(dev, XenbusStateConnected);
1161 		break;
1162 
1163 	case XenbusStateClosing:
1164 		set_backend_state(dev, XenbusStateClosing);
1165 		break;
1166 
1167 	case XenbusStateClosed:
1168 		set_backend_state(dev, XenbusStateClosed);
1169 		if (xenbus_dev_is_online(dev))
1170 			break;
1171 		device_unregister(&dev->dev);
1172 		break;
1173 	case XenbusStateUnknown:
1174 		set_backend_state(dev, XenbusStateClosed);
1175 		device_unregister(&dev->dev);
1176 		break;
1177 
1178 	default:
1179 		xenbus_dev_fatal(dev, -EINVAL, "saw state %d at frontend",
1180 				 frontend_state);
1181 		break;
1182 	}
1183 }
1184 
1185 static int pvcalls_back_remove(struct xenbus_device *dev)
1186 {
1187 	return 0;
1188 }
1189 
1190 static int pvcalls_back_uevent(struct xenbus_device *xdev,
1191 			       struct kobj_uevent_env *env)
1192 {
1193 	return 0;
1194 }
1195 
1196 static const struct xenbus_device_id pvcalls_back_ids[] = {
1197 	{ "pvcalls" },
1198 	{ "" }
1199 };
1200 
1201 static struct xenbus_driver pvcalls_back_driver = {
1202 	.ids = pvcalls_back_ids,
1203 	.probe = pvcalls_back_probe,
1204 	.remove = pvcalls_back_remove,
1205 	.uevent = pvcalls_back_uevent,
1206 	.otherend_changed = pvcalls_back_changed,
1207 };
1208 
1209 static int __init pvcalls_back_init(void)
1210 {
1211 	int ret;
1212 
1213 	if (!xen_domain())
1214 		return -ENODEV;
1215 
1216 	ret = xenbus_register_backend(&pvcalls_back_driver);
1217 	if (ret < 0)
1218 		return ret;
1219 
1220 	sema_init(&pvcalls_back_global.frontends_lock, 1);
1221 	INIT_LIST_HEAD(&pvcalls_back_global.frontends);
1222 	return 0;
1223 }
1224 module_init(pvcalls_back_init);
1225 
1226 static void __exit pvcalls_back_fin(void)
1227 {
1228 	struct pvcalls_fedata *fedata, *nfedata;
1229 
1230 	down(&pvcalls_back_global.frontends_lock);
1231 	list_for_each_entry_safe(fedata, nfedata,
1232 				 &pvcalls_back_global.frontends, list) {
1233 		backend_disconnect(fedata->dev);
1234 	}
1235 	up(&pvcalls_back_global.frontends_lock);
1236 
1237 	xenbus_unregister_driver(&pvcalls_back_driver);
1238 }
1239 
1240 module_exit(pvcalls_back_fin);
1241 
1242 MODULE_DESCRIPTION("Xen PV Calls backend driver");
1243 MODULE_AUTHOR("Stefano Stabellini <sstabellini@kernel.org>");
1244 MODULE_LICENSE("GPL");
1245