drivers/xen/pvcalls-back.c

   1 // SPDX-License-Identifier: GPL-2.0-or-later
   2 /*
   3  * (c) 2017 Stefano Stabellini <[email protected]>
   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 proto_accept_arg arg = {
 521                 .flags = O_NONBLOCK,
 522                 .kern = true,
 523         };
 524         struct sock_mapping *map;
 525         struct pvcalls_ioworker *iow;
 526         struct pvcalls_fedata *fedata;
 527         struct socket *sock;
 528         struct xen_pvcalls_response *rsp;
 529         struct xen_pvcalls_request *req;
 530         int notify;
 531         int ret = -EINVAL;
 532         unsigned long flags;
 533
 534         fedata = mappass->fedata;
 535         /*
 536          * __pvcalls_back_accept can race against pvcalls_back_accept.
 537          * We only need to check the value of "cmd" on read. It could be
 538          * done atomically, but to simplify the code on the write side, we
 539          * use a spinlock.
 540          */
 541         spin_lock_irqsave(&mappass->copy_lock, flags);
 542         req = &mappass->reqcopy;
 543         if (req->cmd != PVCALLS_ACCEPT) {
 544                 spin_unlock_irqrestore(&mappass->copy_lock, flags);
 545                 return;
 546         }
 547         spin_unlock_irqrestore(&mappass->copy_lock, flags);
 548
 549         sock = sock_alloc();
 550         if (sock == NULL)
 551                 goto out_error;
 552         sock->type = mappass->sock->type;
 553         sock->ops = mappass->sock->ops;
 554
 555         ret = inet_accept(mappass->sock, sock, &arg);
 556         if (ret == -EAGAIN) {
 557                 sock_release(sock);
 558                 return;
 559         }
 560
 561         map = pvcalls_new_active_socket(fedata,
 562                                         req->u.accept.id_new,
 563                                         req->u.accept.ref,
 564                                         req->u.accept.evtchn,
 565                                         sock);
 566         if (!map) {
 567                 ret = -EFAULT;
 568                 goto out_error;
 569         }
 570
 571         map->sockpass = mappass;
 572         iow = &map->ioworker;
 573         atomic_inc(&map->read);
 574         atomic_inc(&map->io);
 575         queue_work(iow->wq, &iow->register_work);
 576
 577 out_error:
 578         rsp = RING_GET_RESPONSE(&fedata->ring, fedata->ring.rsp_prod_pvt++);
 579         rsp->req_id = req->req_id;
 580         rsp->cmd = req->cmd;
 581         rsp->u.accept.id = req->u.accept.id;
 582         rsp->ret = ret;
 583         RING_PUSH_RESPONSES_AND_CHECK_NOTIFY(&fedata->ring, notify);
 584         if (notify)
 585                 notify_remote_via_irq(fedata->irq);
 586
 587         mappass->reqcopy.cmd = 0;
 588 }
 589
 590 static void pvcalls_pass_sk_data_ready(struct sock *sock)
 591 {
 592         struct sockpass_mapping *mappass = sock->sk_user_data;
 593         struct pvcalls_fedata *fedata;
 594         struct xen_pvcalls_response *rsp;
 595         unsigned long flags;
 596         int notify;
 597
 598         trace_sk_data_ready(sock);
 599
 600         if (mappass == NULL)
 601                 return;
 602
 603         fedata = mappass->fedata;
 604         spin_lock_irqsave(&mappass->copy_lock, flags);
 605         if (mappass->reqcopy.cmd == PVCALLS_POLL) {
 606                 rsp = RING_GET_RESPONSE(&fedata->ring,
 607                                         fedata->ring.rsp_prod_pvt++);
 608                 rsp->req_id = mappass->reqcopy.req_id;
 609                 rsp->u.poll.id = mappass->reqcopy.u.poll.id;
 610                 rsp->cmd = mappass->reqcopy.cmd;
 611                 rsp->ret = 0;
 612
 613                 mappass->reqcopy.cmd = 0;
 614                 spin_unlock_irqrestore(&mappass->copy_lock, flags);
 615
 616                 RING_PUSH_RESPONSES_AND_CHECK_NOTIFY(&fedata->ring, notify);
 617                 if (notify)
 618                         notify_remote_via_irq(mappass->fedata->irq);
 619         } else {
 620                 spin_unlock_irqrestore(&mappass->copy_lock, flags);
 621                 queue_work(mappass->wq, &mappass->register_work);
 622         }
 623 }
 624
 625 static int pvcalls_back_bind(struct xenbus_device *dev,
 626                              struct xen_pvcalls_request *req)
 627 {
 628         struct pvcalls_fedata *fedata;
 629         int ret;
 630         struct sockpass_mapping *map;
 631         struct xen_pvcalls_response *rsp;
 632
 633         fedata = dev_get_drvdata(&dev->dev);
 634
 635         map = kzalloc(sizeof(*map), GFP_KERNEL);
 636         if (map == NULL) {
 637                 ret = -ENOMEM;
 638                 goto out;
 639         }
 640
 641         INIT_WORK(&map->register_work, __pvcalls_back_accept);
 642         spin_lock_init(&map->copy_lock);
 643         map->wq = alloc_ordered_workqueue("pvcalls_wq", 0);
 644         if (!map->wq) {
 645                 ret = -ENOMEM;
 646                 goto out;
 647         }
 648
 649         ret = sock_create(AF_INET, SOCK_STREAM, 0, &map->sock);
 650         if (ret < 0)
 651                 goto out;
 652
 653         ret = inet_bind(map->sock, (struct sockaddr *)&req->u.bind.addr,
 654                         req->u.bind.len);
 655         if (ret < 0)
 656                 goto out;
 657
 658         map->fedata = fedata;
 659         map->id = req->u.bind.id;
 660
 661         down(&fedata->socket_lock);
 662         ret = radix_tree_insert(&fedata->socketpass_mappings, map->id,
 663                                 map);
 664         up(&fedata->socket_lock);
 665         if (ret)
 666                 goto out;
 667
 668         write_lock_bh(&map->sock->sk->sk_callback_lock);
 669         map->saved_data_ready = map->sock->sk->sk_data_ready;
 670         map->sock->sk->sk_user_data = map;
 671         map->sock->sk->sk_data_ready = pvcalls_pass_sk_data_ready;
 672         write_unlock_bh(&map->sock->sk->sk_callback_lock);
 673
 674 out:
 675         if (ret) {
 676                 if (map && map->sock)
 677                         sock_release(map->sock);
 678                 if (map && map->wq)
 679                         destroy_workqueue(map->wq);
 680                 kfree(map);
 681         }
 682         rsp = RING_GET_RESPONSE(&fedata->ring, fedata->ring.rsp_prod_pvt++);
 683         rsp->req_id = req->req_id;
 684         rsp->cmd = req->cmd;
 685         rsp->u.bind.id = req->u.bind.id;
 686         rsp->ret = ret;
 687         return 0;
 688 }
 689
 690 static int pvcalls_back_listen(struct xenbus_device *dev,
 691                                struct xen_pvcalls_request *req)
 692 {
 693         struct pvcalls_fedata *fedata;
 694         int ret = -EINVAL;
 695         struct sockpass_mapping *map;
 696         struct xen_pvcalls_response *rsp;
 697
 698         fedata = dev_get_drvdata(&dev->dev);
 699
 700         down(&fedata->socket_lock);
 701         map = radix_tree_lookup(&fedata->socketpass_mappings, req->u.listen.id);
 702         up(&fedata->socket_lock);
 703         if (map == NULL)
 704                 goto out;
 705
 706         ret = inet_listen(map->sock, req->u.listen.backlog);
 707
 708 out:
 709         rsp = RING_GET_RESPONSE(&fedata->ring, fedata->ring.rsp_prod_pvt++);
 710         rsp->req_id = req->req_id;
 711         rsp->cmd = req->cmd;
 712         rsp->u.listen.id = req->u.listen.id;
 713         rsp->ret = ret;
 714         return 0;
 715 }
 716
 717 static int pvcalls_back_accept(struct xenbus_device *dev,
 718                                struct xen_pvcalls_request *req)
 719 {
 720         struct pvcalls_fedata *fedata;
 721         struct sockpass_mapping *mappass;
 722         int ret = -EINVAL;
 723         struct xen_pvcalls_response *rsp;
 724         unsigned long flags;
 725
 726         fedata = dev_get_drvdata(&dev->dev);
 727
 728         down(&fedata->socket_lock);
 729         mappass = radix_tree_lookup(&fedata->socketpass_mappings,
 730                 req->u.accept.id);
 731         up(&fedata->socket_lock);
 732         if (mappass == NULL)
 733                 goto out_error;
 734
 735         /*
 736          * Limitation of the current implementation: only support one
 737          * concurrent accept or poll call on one socket.
 738          */
 739         spin_lock_irqsave(&mappass->copy_lock, flags);
 740         if (mappass->reqcopy.cmd != 0) {
 741                 spin_unlock_irqrestore(&mappass->copy_lock, flags);
 742                 ret = -EINTR;
 743                 goto out_error;
 744         }
 745
 746         mappass->reqcopy = *req;
 747         spin_unlock_irqrestore(&mappass->copy_lock, flags);
 748         queue_work(mappass->wq, &mappass->register_work);
 749
 750         /* Tell the caller we don't need to send back a notification yet */
 751         return -1;
 752
 753 out_error:
 754         rsp = RING_GET_RESPONSE(&fedata->ring, fedata->ring.rsp_prod_pvt++);
 755         rsp->req_id = req->req_id;
 756         rsp->cmd = req->cmd;
 757         rsp->u.accept.id = req->u.accept.id;
 758         rsp->ret = ret;
 759         return 0;
 760 }
 761
 762 static int pvcalls_back_poll(struct xenbus_device *dev,
 763                              struct xen_pvcalls_request *req)
 764 {
 765         struct pvcalls_fedata *fedata;
 766         struct sockpass_mapping *mappass;
 767         struct xen_pvcalls_response *rsp;
 768         struct inet_connection_sock *icsk;
 769         struct request_sock_queue *queue;
 770         unsigned long flags;
 771         int ret;
 772         bool data;
 773
 774         fedata = dev_get_drvdata(&dev->dev);
 775
 776         down(&fedata->socket_lock);
 777         mappass = radix_tree_lookup(&fedata->socketpass_mappings,
 778                                     req->u.poll.id);
 779         up(&fedata->socket_lock);
 780         if (mappass == NULL)
 781                 return -EINVAL;
 782
 783         /*
 784          * Limitation of the current implementation: only support one
 785          * concurrent accept or poll call on one socket.
 786          */
 787         spin_lock_irqsave(&mappass->copy_lock, flags);
 788         if (mappass->reqcopy.cmd != 0) {
 789                 ret = -EINTR;
 790                 goto out;
 791         }
 792
 793         mappass->reqcopy = *req;
 794         icsk = inet_csk(mappass->sock->sk);
 795         queue = &icsk->icsk_accept_queue;
 796         data = READ_ONCE(queue->rskq_accept_head) != NULL;
 797         if (data) {
 798                 mappass->reqcopy.cmd = 0;
 799                 ret = 0;
 800                 goto out;
 801         }
 802         spin_unlock_irqrestore(&mappass->copy_lock, flags);
 803
 804         /* Tell the caller we don't need to send back a notification yet */
 805         return -1;
 806
 807 out:
 808         spin_unlock_irqrestore(&mappass->copy_lock, flags);
 809
 810         rsp = RING_GET_RESPONSE(&fedata->ring, fedata->ring.rsp_prod_pvt++);
 811         rsp->req_id = req->req_id;
 812         rsp->cmd = req->cmd;
 813         rsp->u.poll.id = req->u.poll.id;
 814         rsp->ret = ret;
 815         return 0;
 816 }
 817
 818 static int pvcalls_back_handle_cmd(struct xenbus_device *dev,
 819                                    struct xen_pvcalls_request *req)
 820 {
 821         int ret = 0;
 822
 823         switch (req->cmd) {
 824         case PVCALLS_SOCKET:
 825                 ret = pvcalls_back_socket(dev, req);
 826                 break;
 827         case PVCALLS_CONNECT:
 828                 ret = pvcalls_back_connect(dev, req);
 829                 break;
 830         case PVCALLS_RELEASE:
 831                 ret = pvcalls_back_release(dev, req);
 832                 break;
 833         case PVCALLS_BIND:
 834                 ret = pvcalls_back_bind(dev, req);
 835                 break;
 836         case PVCALLS_LISTEN:
 837                 ret = pvcalls_back_listen(dev, req);
 838                 break;
 839         case PVCALLS_ACCEPT:
 840                 ret = pvcalls_back_accept(dev, req);
 841                 break;
 842         case PVCALLS_POLL:
 843                 ret = pvcalls_back_poll(dev, req);
 844                 break;
 845         default:
 846         {
 847                 struct pvcalls_fedata *fedata;
 848                 struct xen_pvcalls_response *rsp;
 849
 850                 fedata = dev_get_drvdata(&dev->dev);
 851                 rsp = RING_GET_RESPONSE(
 852                                 &fedata->ring, fedata->ring.rsp_prod_pvt++);
 853                 rsp->req_id = req->req_id;
 854                 rsp->cmd = req->cmd;
 855                 rsp->ret = -ENOTSUPP;
 856                 break;
 857         }
 858         }
 859         return ret;
 860 }
 861
 862 static void pvcalls_back_work(struct pvcalls_fedata *fedata)
 863 {
 864         int notify, notify_all = 0, more = 1;
 865         struct xen_pvcalls_request req;
 866         struct xenbus_device *dev = fedata->dev;
 867
 868         while (more) {
 869                 while (RING_HAS_UNCONSUMED_REQUESTS(&fedata->ring)) {
 870                         RING_COPY_REQUEST(&fedata->ring,
 871                                           fedata->ring.req_cons++,
 872                                           &req);
 873
 874                         if (!pvcalls_back_handle_cmd(dev, &req)) {
 875                                 RING_PUSH_RESPONSES_AND_CHECK_NOTIFY(
 876                                         &fedata->ring, notify);
 877                                 notify_all += notify;
 878                         }
 879                 }
 880
 881                 if (notify_all) {
 882                         notify_remote_via_irq(fedata->irq);
 883                         notify_all = 0;
 884                 }
 885
 886                 RING_FINAL_CHECK_FOR_REQUESTS(&fedata->ring, more);
 887         }
 888 }
 889
 890 static irqreturn_t pvcalls_back_event(int irq, void *dev_id)
 891 {
 892         struct xenbus_device *dev = dev_id;
 893         struct pvcalls_fedata *fedata = NULL;
 894         unsigned int eoi_flags = XEN_EOI_FLAG_SPURIOUS;
 895
 896         if (dev) {
 897                 fedata = dev_get_drvdata(&dev->dev);
 898                 if (fedata) {
 899                         pvcalls_back_work(fedata);
 900                         eoi_flags = 0;
 901                 }
 902         }
 903
 904         xen_irq_lateeoi(irq, eoi_flags);
 905
 906         return IRQ_HANDLED;
 907 }
 908
 909 static irqreturn_t pvcalls_back_conn_event(int irq, void *sock_map)
 910 {
 911         struct sock_mapping *map = sock_map;
 912         struct pvcalls_ioworker *iow;
 913
 914         if (map == NULL || map->sock == NULL || map->sock->sk == NULL ||
 915                 map->sock->sk->sk_user_data != map) {
 916                 xen_irq_lateeoi(irq, 0);
 917                 return IRQ_HANDLED;
 918         }
 919
 920         iow = &map->ioworker;
 921
 922         atomic_inc(&map->write);
 923         atomic_inc(&map->eoi);
 924         atomic_inc(&map->io);
 925         queue_work(iow->wq, &iow->register_work);
 926
 927         return IRQ_HANDLED;
 928 }
 929
 930 static int backend_connect(struct xenbus_device *dev)
 931 {
 932         int err;
 933         evtchn_port_t evtchn;
 934         grant_ref_t ring_ref;
 935         struct pvcalls_fedata *fedata = NULL;
 936
 937         fedata = kzalloc(sizeof(struct pvcalls_fedata), GFP_KERNEL);
 938         if (!fedata)
 939                 return -ENOMEM;
 940
 941         fedata->irq = -1;
 942         err = xenbus_scanf(XBT_NIL, dev->otherend, "port", "%u",
 943                            &evtchn);
 944         if (err != 1) {
 945                 err = -EINVAL;
 946                 xenbus_dev_fatal(dev, err, "reading %s/event-channel",
 947                                  dev->otherend);
 948                 goto error;
 949         }
 950
 951         err = xenbus_scanf(XBT_NIL, dev->otherend, "ring-ref", "%u", &ring_ref);
 952         if (err != 1) {
 953                 err = -EINVAL;
 954                 xenbus_dev_fatal(dev, err, "reading %s/ring-ref",
 955                                  dev->otherend);
 956                 goto error;
 957         }
 958
 959         err = bind_interdomain_evtchn_to_irq_lateeoi(dev, evtchn);
 960         if (err < 0)
 961                 goto error;
 962         fedata->irq = err;
 963
 964         err = request_threaded_irq(fedata->irq, NULL, pvcalls_back_event,
 965                                    IRQF_ONESHOT, "pvcalls-back", dev);
 966         if (err < 0)
 967                 goto error;
 968
 969         err = xenbus_map_ring_valloc(dev, &ring_ref, 1,
 970                                      (void **)&fedata->sring);
 971         if (err < 0)
 972                 goto error;
 973
 974         BACK_RING_INIT(&fedata->ring, fedata->sring, XEN_PAGE_SIZE * 1);
 975         fedata->dev = dev;
 976
 977         INIT_LIST_HEAD(&fedata->socket_mappings);
 978         INIT_RADIX_TREE(&fedata->socketpass_mappings, GFP_KERNEL);
 979         sema_init(&fedata->socket_lock, 1);
 980         dev_set_drvdata(&dev->dev, fedata);
 981
 982         down(&pvcalls_back_global.frontends_lock);
 983         list_add_tail(&fedata->list, &pvcalls_back_global.frontends);
 984         up(&pvcalls_back_global.frontends_lock);
 985
 986         return 0;
 987
 988  error:
 989         if (fedata->irq >= 0)
 990                 unbind_from_irqhandler(fedata->irq, dev);
 991         if (fedata->sring != NULL)
 992                 xenbus_unmap_ring_vfree(dev, fedata->sring);
 993         kfree(fedata);
 994         return err;
 995 }
 996
 997 static int backend_disconnect(struct xenbus_device *dev)
 998 {
 999         struct pvcalls_fedata *fedata;
1000         struct sock_mapping *map, *n;
1001         struct sockpass_mapping *mappass;
1002         struct radix_tree_iter iter;
1003         void **slot;
1004
1005
1006         fedata = dev_get_drvdata(&dev->dev);
1007
1008         down(&fedata->socket_lock);
1009         list_for_each_entry_safe(map, n, &fedata->socket_mappings, list) {
1010                 list_del(&map->list);
1011                 pvcalls_back_release_active(dev, fedata, map);
1012         }
1013
1014         radix_tree_for_each_slot(slot, &fedata->socketpass_mappings, &iter, 0) {
1015                 mappass = radix_tree_deref_slot(slot);
1016                 if (!mappass)
1017                         continue;
1018                 if (radix_tree_exception(mappass)) {
1019                         if (radix_tree_deref_retry(mappass))
1020                                 slot = radix_tree_iter_retry(&iter);
1021                 } else {
1022                         radix_tree_delete(&fedata->socketpass_mappings,
1023                                           mappass->id);
1024                         pvcalls_back_release_passive(dev, fedata, mappass);
1025                 }
1026         }
1027         up(&fedata->socket_lock);
1028
1029         unbind_from_irqhandler(fedata->irq, dev);
1030         xenbus_unmap_ring_vfree(dev, fedata->sring);
1031
1032         list_del(&fedata->list);
1033         kfree(fedata);
1034         dev_set_drvdata(&dev->dev, NULL);
1035
1036         return 0;
1037 }
1038
1039 static int pvcalls_back_probe(struct xenbus_device *dev,
1040                               const struct xenbus_device_id *id)
1041 {
1042         int err, abort;
1043         struct xenbus_transaction xbt;
1044
1045 again:
1046         abort = 1;
1047
1048         err = xenbus_transaction_start(&xbt);
1049         if (err) {
1050                 pr_warn("%s cannot create xenstore transaction\n", __func__);
1051                 return err;
1052         }
1053
1054         err = xenbus_printf(xbt, dev->nodename, "versions", "%s",
1055                             PVCALLS_VERSIONS);
1056         if (err) {
1057                 pr_warn("%s write out 'versions' failed\n", __func__);
1058                 goto abort;
1059         }
1060
1061         err = xenbus_printf(xbt, dev->nodename, "max-page-order", "%u",
1062                             MAX_RING_ORDER);
1063         if (err) {
1064                 pr_warn("%s write out 'max-page-order' failed\n", __func__);
1065                 goto abort;
1066         }
1067
1068         err = xenbus_printf(xbt, dev->nodename, "function-calls",
1069                             XENBUS_FUNCTIONS_CALLS);
1070         if (err) {
1071                 pr_warn("%s write out 'function-calls' failed\n", __func__);
1072                 goto abort;
1073         }
1074
1075         abort = 0;
1076 abort:
1077         err = xenbus_transaction_end(xbt, abort);
1078         if (err) {
1079                 if (err == -EAGAIN && !abort)
1080                         goto again;
1081                 pr_warn("%s cannot complete xenstore transaction\n", __func__);
1082                 return err;
1083         }
1084
1085         if (abort)
1086                 return -EFAULT;
1087
1088         xenbus_switch_state(dev, XenbusStateInitWait);
1089
1090         return 0;
1091 }
1092
1093 static void set_backend_state(struct xenbus_device *dev,
1094                               enum xenbus_state state)
1095 {
1096         while (dev->state != state) {
1097                 switch (dev->state) {
1098                 case XenbusStateClosed:
1099                         switch (state) {
1100                         case XenbusStateInitWait:
1101                         case XenbusStateConnected:
1102                                 xenbus_switch_state(dev, XenbusStateInitWait);
1103                                 break;
1104                         case XenbusStateClosing:
1105                                 xenbus_switch_state(dev, XenbusStateClosing);
1106                                 break;
1107                         default:
1108                                 WARN_ON(1);
1109                         }
1110                         break;
1111                 case XenbusStateInitWait:
1112                 case XenbusStateInitialised:
1113                         switch (state) {
1114                         case XenbusStateConnected:
1115                                 if (backend_connect(dev))
1116                                         return;
1117                                 xenbus_switch_state(dev, XenbusStateConnected);
1118                                 break;
1119                         case XenbusStateClosing:
1120                         case XenbusStateClosed:
1121                                 xenbus_switch_state(dev, XenbusStateClosing);
1122                                 break;
1123                         default:
1124                                 WARN_ON(1);
1125                         }
1126                         break;
1127                 case XenbusStateConnected:
1128                         switch (state) {
1129                         case XenbusStateInitWait:
1130                         case XenbusStateClosing:
1131                         case XenbusStateClosed:
1132                                 down(&pvcalls_back_global.frontends_lock);
1133                                 backend_disconnect(dev);
1134                                 up(&pvcalls_back_global.frontends_lock);
1135                                 xenbus_switch_state(dev, XenbusStateClosing);
1136                                 break;
1137                         default:
1138                                 WARN_ON(1);
1139                         }
1140                         break;
1141                 case XenbusStateClosing:
1142                         switch (state) {
1143                         case XenbusStateInitWait:
1144                         case XenbusStateConnected:
1145                         case XenbusStateClosed:
1146                                 xenbus_switch_state(dev, XenbusStateClosed);
1147                                 break;
1148                         default:
1149                                 WARN_ON(1);
1150                         }
1151                         break;
1152                 default:
1153                         WARN_ON(1);
1154                 }
1155         }
1156 }
1157
1158 static void pvcalls_back_changed(struct xenbus_device *dev,
1159                                  enum xenbus_state frontend_state)
1160 {
1161         switch (frontend_state) {
1162         case XenbusStateInitialising:
1163                 set_backend_state(dev, XenbusStateInitWait);
1164                 break;
1165
1166         case XenbusStateInitialised:
1167         case XenbusStateConnected:
1168                 set_backend_state(dev, XenbusStateConnected);
1169                 break;
1170
1171         case XenbusStateClosing:
1172                 set_backend_state(dev, XenbusStateClosing);
1173                 break;
1174
1175         case XenbusStateClosed:
1176                 set_backend_state(dev, XenbusStateClosed);
1177                 if (xenbus_dev_is_online(dev))
1178                         break;
1179                 device_unregister(&dev->dev);
1180                 break;
1181         case XenbusStateUnknown:
1182                 set_backend_state(dev, XenbusStateClosed);
1183                 device_unregister(&dev->dev);
1184                 break;
1185
1186         default:
1187                 xenbus_dev_fatal(dev, -EINVAL, "saw state %d at frontend",
1188                                  frontend_state);
1189                 break;
1190         }
1191 }
1192
1193 static void pvcalls_back_remove(struct xenbus_device *dev)
1194 {
1195 }
1196
1197 static int pvcalls_back_uevent(const struct xenbus_device *xdev,
1198                                struct kobj_uevent_env *env)
1199 {
1200         return 0;
1201 }
1202
1203 static const struct xenbus_device_id pvcalls_back_ids[] = {
1204         { "pvcalls" },
1205         { "" }
1206 };
1207
1208 static struct xenbus_driver pvcalls_back_driver = {
1209         .ids = pvcalls_back_ids,
1210         .probe = pvcalls_back_probe,
1211         .remove = pvcalls_back_remove,
1212         .uevent = pvcalls_back_uevent,
1213         .otherend_changed = pvcalls_back_changed,
1214 };
1215
1216 static int __init pvcalls_back_init(void)
1217 {
1218         int ret;
1219
1220         if (!xen_domain())
1221                 return -ENODEV;
1222
1223         ret = xenbus_register_backend(&pvcalls_back_driver);
1224         if (ret < 0)
1225                 return ret;
1226
1227         sema_init(&pvcalls_back_global.frontends_lock, 1);
1228         INIT_LIST_HEAD(&pvcalls_back_global.frontends);
1229         return 0;
1230 }
1231 module_init(pvcalls_back_init);
1232
1233 static void __exit pvcalls_back_fin(void)
1234 {
1235         struct pvcalls_fedata *fedata, *nfedata;
1236
1237         down(&pvcalls_back_global.frontends_lock);
1238         list_for_each_entry_safe(fedata, nfedata,
1239                                  &pvcalls_back_global.frontends, list) {
1240                 backend_disconnect(fedata->dev);
1241         }
1242         up(&pvcalls_back_global.frontends_lock);
1243
1244         xenbus_unregister_driver(&pvcalls_back_driver);
1245 }
1246
1247 module_exit(pvcalls_back_fin);
1248
1249 MODULE_DESCRIPTION("Xen PV Calls backend driver");
1250 MODULE_AUTHOR("Stefano Stabellini <[email protected]>");
1251 MODULE_LICENSE("GPL");