1 // SPDX-License-Identifier: GPL-2.0-or-later 2 /* 3 * (c) 2017 Stefano Stabellini <stefano@aporeto.com> 4 */ 5 6 #include <linux/module.h> 7 #include <linux/net.h> 8 #include <linux/socket.h> 9 10 #include <net/sock.h> 11 12 #include <xen/events.h> 13 #include <xen/grant_table.h> 14 #include <xen/xen.h> 15 #include <xen/xenbus.h> 16 #include <xen/interface/io/pvcalls.h> 17 18 #include "pvcalls-front.h" 19 20 #define PVCALLS_INVALID_ID UINT_MAX 21 #define PVCALLS_RING_ORDER XENBUS_MAX_RING_GRANT_ORDER 22 #define PVCALLS_NR_RSP_PER_RING __CONST_RING_SIZE(xen_pvcalls, XEN_PAGE_SIZE) 23 #define PVCALLS_FRONT_MAX_SPIN 5000 24 25 static struct proto pvcalls_proto = { 26 .name = "PVCalls", 27 .owner = THIS_MODULE, 28 .obj_size = sizeof(struct sock), 29 }; 30 31 struct pvcalls_bedata { 32 struct xen_pvcalls_front_ring ring; 33 grant_ref_t ref; 34 int irq; 35 36 struct list_head socket_mappings; 37 spinlock_t socket_lock; 38 39 wait_queue_head_t inflight_req; 40 struct xen_pvcalls_response rsp[PVCALLS_NR_RSP_PER_RING]; 41 }; 42 /* Only one front/back connection supported. */ 43 static struct xenbus_device *pvcalls_front_dev; 44 static atomic_t pvcalls_refcount; 45 46 /* first increment refcount, then proceed */ 47 #define pvcalls_enter() { \ 48 atomic_inc(&pvcalls_refcount); \ 49 } 50 51 /* first complete other operations, then decrement refcount */ 52 #define pvcalls_exit() { \ 53 atomic_dec(&pvcalls_refcount); \ 54 } 55 56 struct sock_mapping { 57 bool active_socket; 58 struct list_head list; 59 struct socket *sock; 60 atomic_t refcount; 61 union { 62 struct { 63 int irq; 64 grant_ref_t ref; 65 struct pvcalls_data_intf *ring; 66 struct pvcalls_data data; 67 struct mutex in_mutex; 68 struct mutex out_mutex; 69 70 wait_queue_head_t inflight_conn_req; 71 } active; 72 struct { 73 /* 74 * Socket status, needs to be 64-bit aligned due to the 75 * test_and_* functions which have this requirement on arm64. 76 */ 77 #define PVCALLS_STATUS_UNINITALIZED 0 78 #define PVCALLS_STATUS_BIND 1 79 #define PVCALLS_STATUS_LISTEN 2 80 uint8_t status __attribute__((aligned(8))); 81 /* 82 * Internal state-machine flags. 83 * Only one accept operation can be inflight for a socket. 84 * Only one poll operation can be inflight for a given socket. 85 * flags needs to be 64-bit aligned due to the test_and_* 86 * functions which have this requirement on arm64. 87 */ 88 #define PVCALLS_FLAG_ACCEPT_INFLIGHT 0 89 #define PVCALLS_FLAG_POLL_INFLIGHT 1 90 #define PVCALLS_FLAG_POLL_RET 2 91 uint8_t flags __attribute__((aligned(8))); 92 uint32_t inflight_req_id; 93 struct sock_mapping *accept_map; 94 wait_queue_head_t inflight_accept_req; 95 } passive; 96 }; 97 }; 98 99 static inline struct sock_mapping *pvcalls_enter_sock(struct socket *sock) 100 { 101 struct sock_mapping *map; 102 103 if (!pvcalls_front_dev || 104 dev_get_drvdata(&pvcalls_front_dev->dev) == NULL) 105 return ERR_PTR(-ENOTCONN); 106 107 map = (struct sock_mapping *)sock->sk->sk_send_head; 108 if (map == NULL) 109 return ERR_PTR(-ENOTSOCK); 110 111 pvcalls_enter(); 112 atomic_inc(&map->refcount); 113 return map; 114 } 115 116 static inline void pvcalls_exit_sock(struct socket *sock) 117 { 118 struct sock_mapping *map; 119 120 map = (struct sock_mapping *)sock->sk->sk_send_head; 121 atomic_dec(&map->refcount); 122 pvcalls_exit(); 123 } 124 125 static inline int get_request(struct pvcalls_bedata *bedata, int *req_id) 126 { 127 *req_id = bedata->ring.req_prod_pvt & (RING_SIZE(&bedata->ring) - 1); 128 if (RING_FULL(&bedata->ring) || 129 bedata->rsp[*req_id].req_id != PVCALLS_INVALID_ID) 130 return -EAGAIN; 131 return 0; 132 } 133 134 static bool pvcalls_front_write_todo(struct sock_mapping *map) 135 { 136 struct pvcalls_data_intf *intf = map->active.ring; 137 RING_IDX cons, prod, size = XEN_FLEX_RING_SIZE(PVCALLS_RING_ORDER); 138 int32_t error; 139 140 error = intf->out_error; 141 if (error == -ENOTCONN) 142 return false; 143 if (error != 0) 144 return true; 145 146 cons = intf->out_cons; 147 prod = intf->out_prod; 148 return !!(size - pvcalls_queued(prod, cons, size)); 149 } 150 151 static bool pvcalls_front_read_todo(struct sock_mapping *map) 152 { 153 struct pvcalls_data_intf *intf = map->active.ring; 154 RING_IDX cons, prod; 155 int32_t error; 156 157 cons = intf->in_cons; 158 prod = intf->in_prod; 159 error = intf->in_error; 160 return (error != 0 || 161 pvcalls_queued(prod, cons, 162 XEN_FLEX_RING_SIZE(PVCALLS_RING_ORDER)) != 0); 163 } 164 165 static irqreturn_t pvcalls_front_event_handler(int irq, void *dev_id) 166 { 167 struct xenbus_device *dev = dev_id; 168 struct pvcalls_bedata *bedata; 169 struct xen_pvcalls_response *rsp; 170 uint8_t *src, *dst; 171 int req_id = 0, more = 0, done = 0; 172 173 if (dev == NULL) 174 return IRQ_HANDLED; 175 176 pvcalls_enter(); 177 bedata = dev_get_drvdata(&dev->dev); 178 if (bedata == NULL) { 179 pvcalls_exit(); 180 return IRQ_HANDLED; 181 } 182 183 again: 184 while (RING_HAS_UNCONSUMED_RESPONSES(&bedata->ring)) { 185 rsp = RING_GET_RESPONSE(&bedata->ring, bedata->ring.rsp_cons); 186 187 req_id = rsp->req_id; 188 if (rsp->cmd == PVCALLS_POLL) { 189 struct sock_mapping *map = (struct sock_mapping *)(uintptr_t) 190 rsp->u.poll.id; 191 192 clear_bit(PVCALLS_FLAG_POLL_INFLIGHT, 193 (void *)&map->passive.flags); 194 /* 195 * clear INFLIGHT, then set RET. It pairs with 196 * the checks at the beginning of 197 * pvcalls_front_poll_passive. 198 */ 199 smp_wmb(); 200 set_bit(PVCALLS_FLAG_POLL_RET, 201 (void *)&map->passive.flags); 202 } else { 203 dst = (uint8_t *)&bedata->rsp[req_id] + 204 sizeof(rsp->req_id); 205 src = (uint8_t *)rsp + sizeof(rsp->req_id); 206 memcpy(dst, src, sizeof(*rsp) - sizeof(rsp->req_id)); 207 /* 208 * First copy the rest of the data, then req_id. It is 209 * paired with the barrier when accessing bedata->rsp. 210 */ 211 smp_wmb(); 212 bedata->rsp[req_id].req_id = req_id; 213 } 214 215 done = 1; 216 bedata->ring.rsp_cons++; 217 } 218 219 RING_FINAL_CHECK_FOR_RESPONSES(&bedata->ring, more); 220 if (more) 221 goto again; 222 if (done) 223 wake_up(&bedata->inflight_req); 224 pvcalls_exit(); 225 return IRQ_HANDLED; 226 } 227 228 static void pvcalls_front_free_map(struct pvcalls_bedata *bedata, 229 struct sock_mapping *map) 230 { 231 int i; 232 233 unbind_from_irqhandler(map->active.irq, map); 234 235 spin_lock(&bedata->socket_lock); 236 if (!list_empty(&map->list)) 237 list_del_init(&map->list); 238 spin_unlock(&bedata->socket_lock); 239 240 for (i = 0; i < (1 << PVCALLS_RING_ORDER); i++) 241 gnttab_end_foreign_access(map->active.ring->ref[i], 0, 0); 242 gnttab_end_foreign_access(map->active.ref, 0, 0); 243 free_page((unsigned long)map->active.ring); 244 245 kfree(map); 246 } 247 248 static irqreturn_t pvcalls_front_conn_handler(int irq, void *sock_map) 249 { 250 struct sock_mapping *map = sock_map; 251 252 if (map == NULL) 253 return IRQ_HANDLED; 254 255 wake_up_interruptible(&map->active.inflight_conn_req); 256 257 return IRQ_HANDLED; 258 } 259 260 int pvcalls_front_socket(struct socket *sock) 261 { 262 struct pvcalls_bedata *bedata; 263 struct sock_mapping *map = NULL; 264 struct xen_pvcalls_request *req; 265 int notify, req_id, ret; 266 267 /* 268 * PVCalls only supports domain AF_INET, 269 * type SOCK_STREAM and protocol 0 sockets for now. 270 * 271 * Check socket type here, AF_INET and protocol checks are done 272 * by the caller. 273 */ 274 if (sock->type != SOCK_STREAM) 275 return -EOPNOTSUPP; 276 277 pvcalls_enter(); 278 if (!pvcalls_front_dev) { 279 pvcalls_exit(); 280 return -EACCES; 281 } 282 bedata = dev_get_drvdata(&pvcalls_front_dev->dev); 283 284 map = kzalloc(sizeof(*map), GFP_KERNEL); 285 if (map == NULL) { 286 pvcalls_exit(); 287 return -ENOMEM; 288 } 289 290 spin_lock(&bedata->socket_lock); 291 292 ret = get_request(bedata, &req_id); 293 if (ret < 0) { 294 kfree(map); 295 spin_unlock(&bedata->socket_lock); 296 pvcalls_exit(); 297 return ret; 298 } 299 300 /* 301 * sock->sk->sk_send_head is not used for ip sockets: reuse the 302 * field to store a pointer to the struct sock_mapping 303 * corresponding to the socket. This way, we can easily get the 304 * struct sock_mapping from the struct socket. 305 */ 306 sock->sk->sk_send_head = (void *)map; 307 list_add_tail(&map->list, &bedata->socket_mappings); 308 309 req = RING_GET_REQUEST(&bedata->ring, req_id); 310 req->req_id = req_id; 311 req->cmd = PVCALLS_SOCKET; 312 req->u.socket.id = (uintptr_t) map; 313 req->u.socket.domain = AF_INET; 314 req->u.socket.type = SOCK_STREAM; 315 req->u.socket.protocol = IPPROTO_IP; 316 317 bedata->ring.req_prod_pvt++; 318 RING_PUSH_REQUESTS_AND_CHECK_NOTIFY(&bedata->ring, notify); 319 spin_unlock(&bedata->socket_lock); 320 if (notify) 321 notify_remote_via_irq(bedata->irq); 322 323 wait_event(bedata->inflight_req, 324 READ_ONCE(bedata->rsp[req_id].req_id) == req_id); 325 326 /* read req_id, then the content */ 327 smp_rmb(); 328 ret = bedata->rsp[req_id].ret; 329 bedata->rsp[req_id].req_id = PVCALLS_INVALID_ID; 330 331 pvcalls_exit(); 332 return ret; 333 } 334 335 static void free_active_ring(struct sock_mapping *map) 336 { 337 if (!map->active.ring) 338 return; 339 340 free_pages((unsigned long)map->active.data.in, 341 map->active.ring->ring_order); 342 free_page((unsigned long)map->active.ring); 343 } 344 345 static int alloc_active_ring(struct sock_mapping *map) 346 { 347 void *bytes; 348 349 map->active.ring = (struct pvcalls_data_intf *) 350 get_zeroed_page(GFP_KERNEL); 351 if (!map->active.ring) 352 goto out; 353 354 map->active.ring->ring_order = PVCALLS_RING_ORDER; 355 bytes = (void *)__get_free_pages(GFP_KERNEL | __GFP_ZERO, 356 PVCALLS_RING_ORDER); 357 if (!bytes) 358 goto out; 359 360 map->active.data.in = bytes; 361 map->active.data.out = bytes + 362 XEN_FLEX_RING_SIZE(PVCALLS_RING_ORDER); 363 364 return 0; 365 366 out: 367 free_active_ring(map); 368 return -ENOMEM; 369 } 370 371 static int create_active(struct sock_mapping *map, int *evtchn) 372 { 373 void *bytes; 374 int ret = -ENOMEM, irq = -1, i; 375 376 *evtchn = -1; 377 init_waitqueue_head(&map->active.inflight_conn_req); 378 379 bytes = map->active.data.in; 380 for (i = 0; i < (1 << PVCALLS_RING_ORDER); i++) 381 map->active.ring->ref[i] = gnttab_grant_foreign_access( 382 pvcalls_front_dev->otherend_id, 383 pfn_to_gfn(virt_to_pfn(bytes) + i), 0); 384 385 map->active.ref = gnttab_grant_foreign_access( 386 pvcalls_front_dev->otherend_id, 387 pfn_to_gfn(virt_to_pfn((void *)map->active.ring)), 0); 388 389 ret = xenbus_alloc_evtchn(pvcalls_front_dev, evtchn); 390 if (ret) 391 goto out_error; 392 irq = bind_evtchn_to_irqhandler(*evtchn, pvcalls_front_conn_handler, 393 0, "pvcalls-frontend", map); 394 if (irq < 0) { 395 ret = irq; 396 goto out_error; 397 } 398 399 map->active.irq = irq; 400 map->active_socket = true; 401 mutex_init(&map->active.in_mutex); 402 mutex_init(&map->active.out_mutex); 403 404 return 0; 405 406 out_error: 407 if (*evtchn >= 0) 408 xenbus_free_evtchn(pvcalls_front_dev, *evtchn); 409 return ret; 410 } 411 412 int pvcalls_front_connect(struct socket *sock, struct sockaddr *addr, 413 int addr_len, int flags) 414 { 415 struct pvcalls_bedata *bedata; 416 struct sock_mapping *map = NULL; 417 struct xen_pvcalls_request *req; 418 int notify, req_id, ret, evtchn; 419 420 if (addr->sa_family != AF_INET || sock->type != SOCK_STREAM) 421 return -EOPNOTSUPP; 422 423 map = pvcalls_enter_sock(sock); 424 if (IS_ERR(map)) 425 return PTR_ERR(map); 426 427 bedata = dev_get_drvdata(&pvcalls_front_dev->dev); 428 ret = alloc_active_ring(map); 429 if (ret < 0) { 430 pvcalls_exit_sock(sock); 431 return ret; 432 } 433 434 spin_lock(&bedata->socket_lock); 435 ret = get_request(bedata, &req_id); 436 if (ret < 0) { 437 spin_unlock(&bedata->socket_lock); 438 free_active_ring(map); 439 pvcalls_exit_sock(sock); 440 return ret; 441 } 442 ret = create_active(map, &evtchn); 443 if (ret < 0) { 444 spin_unlock(&bedata->socket_lock); 445 free_active_ring(map); 446 pvcalls_exit_sock(sock); 447 return ret; 448 } 449 450 req = RING_GET_REQUEST(&bedata->ring, req_id); 451 req->req_id = req_id; 452 req->cmd = PVCALLS_CONNECT; 453 req->u.connect.id = (uintptr_t)map; 454 req->u.connect.len = addr_len; 455 req->u.connect.flags = flags; 456 req->u.connect.ref = map->active.ref; 457 req->u.connect.evtchn = evtchn; 458 memcpy(req->u.connect.addr, addr, sizeof(*addr)); 459 460 map->sock = sock; 461 462 bedata->ring.req_prod_pvt++; 463 RING_PUSH_REQUESTS_AND_CHECK_NOTIFY(&bedata->ring, notify); 464 spin_unlock(&bedata->socket_lock); 465 466 if (notify) 467 notify_remote_via_irq(bedata->irq); 468 469 wait_event(bedata->inflight_req, 470 READ_ONCE(bedata->rsp[req_id].req_id) == req_id); 471 472 /* read req_id, then the content */ 473 smp_rmb(); 474 ret = bedata->rsp[req_id].ret; 475 bedata->rsp[req_id].req_id = PVCALLS_INVALID_ID; 476 pvcalls_exit_sock(sock); 477 return ret; 478 } 479 480 static int __write_ring(struct pvcalls_data_intf *intf, 481 struct pvcalls_data *data, 482 struct iov_iter *msg_iter, 483 int len) 484 { 485 RING_IDX cons, prod, size, masked_prod, masked_cons; 486 RING_IDX array_size = XEN_FLEX_RING_SIZE(PVCALLS_RING_ORDER); 487 int32_t error; 488 489 error = intf->out_error; 490 if (error < 0) 491 return error; 492 cons = intf->out_cons; 493 prod = intf->out_prod; 494 /* read indexes before continuing */ 495 virt_mb(); 496 497 size = pvcalls_queued(prod, cons, array_size); 498 if (size > array_size) 499 return -EINVAL; 500 if (size == array_size) 501 return 0; 502 if (len > array_size - size) 503 len = array_size - size; 504 505 masked_prod = pvcalls_mask(prod, array_size); 506 masked_cons = pvcalls_mask(cons, array_size); 507 508 if (masked_prod < masked_cons) { 509 len = copy_from_iter(data->out + masked_prod, len, msg_iter); 510 } else { 511 if (len > array_size - masked_prod) { 512 int ret = copy_from_iter(data->out + masked_prod, 513 array_size - masked_prod, msg_iter); 514 if (ret != array_size - masked_prod) { 515 len = ret; 516 goto out; 517 } 518 len = ret + copy_from_iter(data->out, len - ret, msg_iter); 519 } else { 520 len = copy_from_iter(data->out + masked_prod, len, msg_iter); 521 } 522 } 523 out: 524 /* write to ring before updating pointer */ 525 virt_wmb(); 526 intf->out_prod += len; 527 528 return len; 529 } 530 531 int pvcalls_front_sendmsg(struct socket *sock, struct msghdr *msg, 532 size_t len) 533 { 534 struct pvcalls_bedata *bedata; 535 struct sock_mapping *map; 536 int sent, tot_sent = 0; 537 int count = 0, flags; 538 539 flags = msg->msg_flags; 540 if (flags & (MSG_CONFIRM|MSG_DONTROUTE|MSG_EOR|MSG_OOB)) 541 return -EOPNOTSUPP; 542 543 map = pvcalls_enter_sock(sock); 544 if (IS_ERR(map)) 545 return PTR_ERR(map); 546 bedata = dev_get_drvdata(&pvcalls_front_dev->dev); 547 548 mutex_lock(&map->active.out_mutex); 549 if ((flags & MSG_DONTWAIT) && !pvcalls_front_write_todo(map)) { 550 mutex_unlock(&map->active.out_mutex); 551 pvcalls_exit_sock(sock); 552 return -EAGAIN; 553 } 554 if (len > INT_MAX) 555 len = INT_MAX; 556 557 again: 558 count++; 559 sent = __write_ring(map->active.ring, 560 &map->active.data, &msg->msg_iter, 561 len); 562 if (sent > 0) { 563 len -= sent; 564 tot_sent += sent; 565 notify_remote_via_irq(map->active.irq); 566 } 567 if (sent >= 0 && len > 0 && count < PVCALLS_FRONT_MAX_SPIN) 568 goto again; 569 if (sent < 0) 570 tot_sent = sent; 571 572 mutex_unlock(&map->active.out_mutex); 573 pvcalls_exit_sock(sock); 574 return tot_sent; 575 } 576 577 static int __read_ring(struct pvcalls_data_intf *intf, 578 struct pvcalls_data *data, 579 struct iov_iter *msg_iter, 580 size_t len, int flags) 581 { 582 RING_IDX cons, prod, size, masked_prod, masked_cons; 583 RING_IDX array_size = XEN_FLEX_RING_SIZE(PVCALLS_RING_ORDER); 584 int32_t error; 585 586 cons = intf->in_cons; 587 prod = intf->in_prod; 588 error = intf->in_error; 589 /* get pointers before reading from the ring */ 590 virt_rmb(); 591 592 size = pvcalls_queued(prod, cons, array_size); 593 masked_prod = pvcalls_mask(prod, array_size); 594 masked_cons = pvcalls_mask(cons, array_size); 595 596 if (size == 0) 597 return error ?: size; 598 599 if (len > size) 600 len = size; 601 602 if (masked_prod > masked_cons) { 603 len = copy_to_iter(data->in + masked_cons, len, msg_iter); 604 } else { 605 if (len > (array_size - masked_cons)) { 606 int ret = copy_to_iter(data->in + masked_cons, 607 array_size - masked_cons, msg_iter); 608 if (ret != array_size - masked_cons) { 609 len = ret; 610 goto out; 611 } 612 len = ret + copy_to_iter(data->in, len - ret, msg_iter); 613 } else { 614 len = copy_to_iter(data->in + masked_cons, len, msg_iter); 615 } 616 } 617 out: 618 /* read data from the ring before increasing the index */ 619 virt_mb(); 620 if (!(flags & MSG_PEEK)) 621 intf->in_cons += len; 622 623 return len; 624 } 625 626 int pvcalls_front_recvmsg(struct socket *sock, struct msghdr *msg, size_t len, 627 int flags) 628 { 629 struct pvcalls_bedata *bedata; 630 int ret; 631 struct sock_mapping *map; 632 633 if (flags & (MSG_CMSG_CLOEXEC|MSG_ERRQUEUE|MSG_OOB|MSG_TRUNC)) 634 return -EOPNOTSUPP; 635 636 map = pvcalls_enter_sock(sock); 637 if (IS_ERR(map)) 638 return PTR_ERR(map); 639 bedata = dev_get_drvdata(&pvcalls_front_dev->dev); 640 641 mutex_lock(&map->active.in_mutex); 642 if (len > XEN_FLEX_RING_SIZE(PVCALLS_RING_ORDER)) 643 len = XEN_FLEX_RING_SIZE(PVCALLS_RING_ORDER); 644 645 while (!(flags & MSG_DONTWAIT) && !pvcalls_front_read_todo(map)) { 646 wait_event_interruptible(map->active.inflight_conn_req, 647 pvcalls_front_read_todo(map)); 648 } 649 ret = __read_ring(map->active.ring, &map->active.data, 650 &msg->msg_iter, len, flags); 651 652 if (ret > 0) 653 notify_remote_via_irq(map->active.irq); 654 if (ret == 0) 655 ret = (flags & MSG_DONTWAIT) ? -EAGAIN : 0; 656 if (ret == -ENOTCONN) 657 ret = 0; 658 659 mutex_unlock(&map->active.in_mutex); 660 pvcalls_exit_sock(sock); 661 return ret; 662 } 663 664 int pvcalls_front_bind(struct socket *sock, struct sockaddr *addr, int addr_len) 665 { 666 struct pvcalls_bedata *bedata; 667 struct sock_mapping *map = NULL; 668 struct xen_pvcalls_request *req; 669 int notify, req_id, ret; 670 671 if (addr->sa_family != AF_INET || sock->type != SOCK_STREAM) 672 return -EOPNOTSUPP; 673 674 map = pvcalls_enter_sock(sock); 675 if (IS_ERR(map)) 676 return PTR_ERR(map); 677 bedata = dev_get_drvdata(&pvcalls_front_dev->dev); 678 679 spin_lock(&bedata->socket_lock); 680 ret = get_request(bedata, &req_id); 681 if (ret < 0) { 682 spin_unlock(&bedata->socket_lock); 683 pvcalls_exit_sock(sock); 684 return ret; 685 } 686 req = RING_GET_REQUEST(&bedata->ring, req_id); 687 req->req_id = req_id; 688 map->sock = sock; 689 req->cmd = PVCALLS_BIND; 690 req->u.bind.id = (uintptr_t)map; 691 memcpy(req->u.bind.addr, addr, sizeof(*addr)); 692 req->u.bind.len = addr_len; 693 694 init_waitqueue_head(&map->passive.inflight_accept_req); 695 696 map->active_socket = false; 697 698 bedata->ring.req_prod_pvt++; 699 RING_PUSH_REQUESTS_AND_CHECK_NOTIFY(&bedata->ring, notify); 700 spin_unlock(&bedata->socket_lock); 701 if (notify) 702 notify_remote_via_irq(bedata->irq); 703 704 wait_event(bedata->inflight_req, 705 READ_ONCE(bedata->rsp[req_id].req_id) == req_id); 706 707 /* read req_id, then the content */ 708 smp_rmb(); 709 ret = bedata->rsp[req_id].ret; 710 bedata->rsp[req_id].req_id = PVCALLS_INVALID_ID; 711 712 map->passive.status = PVCALLS_STATUS_BIND; 713 pvcalls_exit_sock(sock); 714 return 0; 715 } 716 717 int pvcalls_front_listen(struct socket *sock, int backlog) 718 { 719 struct pvcalls_bedata *bedata; 720 struct sock_mapping *map; 721 struct xen_pvcalls_request *req; 722 int notify, req_id, ret; 723 724 map = pvcalls_enter_sock(sock); 725 if (IS_ERR(map)) 726 return PTR_ERR(map); 727 bedata = dev_get_drvdata(&pvcalls_front_dev->dev); 728 729 if (map->passive.status != PVCALLS_STATUS_BIND) { 730 pvcalls_exit_sock(sock); 731 return -EOPNOTSUPP; 732 } 733 734 spin_lock(&bedata->socket_lock); 735 ret = get_request(bedata, &req_id); 736 if (ret < 0) { 737 spin_unlock(&bedata->socket_lock); 738 pvcalls_exit_sock(sock); 739 return ret; 740 } 741 req = RING_GET_REQUEST(&bedata->ring, req_id); 742 req->req_id = req_id; 743 req->cmd = PVCALLS_LISTEN; 744 req->u.listen.id = (uintptr_t) map; 745 req->u.listen.backlog = backlog; 746 747 bedata->ring.req_prod_pvt++; 748 RING_PUSH_REQUESTS_AND_CHECK_NOTIFY(&bedata->ring, notify); 749 spin_unlock(&bedata->socket_lock); 750 if (notify) 751 notify_remote_via_irq(bedata->irq); 752 753 wait_event(bedata->inflight_req, 754 READ_ONCE(bedata->rsp[req_id].req_id) == req_id); 755 756 /* read req_id, then the content */ 757 smp_rmb(); 758 ret = bedata->rsp[req_id].ret; 759 bedata->rsp[req_id].req_id = PVCALLS_INVALID_ID; 760 761 map->passive.status = PVCALLS_STATUS_LISTEN; 762 pvcalls_exit_sock(sock); 763 return ret; 764 } 765 766 int pvcalls_front_accept(struct socket *sock, struct socket *newsock, int flags) 767 { 768 struct pvcalls_bedata *bedata; 769 struct sock_mapping *map; 770 struct sock_mapping *map2 = NULL; 771 struct xen_pvcalls_request *req; 772 int notify, req_id, ret, evtchn, nonblock; 773 774 map = pvcalls_enter_sock(sock); 775 if (IS_ERR(map)) 776 return PTR_ERR(map); 777 bedata = dev_get_drvdata(&pvcalls_front_dev->dev); 778 779 if (map->passive.status != PVCALLS_STATUS_LISTEN) { 780 pvcalls_exit_sock(sock); 781 return -EINVAL; 782 } 783 784 nonblock = flags & SOCK_NONBLOCK; 785 /* 786 * Backend only supports 1 inflight accept request, will return 787 * errors for the others 788 */ 789 if (test_and_set_bit(PVCALLS_FLAG_ACCEPT_INFLIGHT, 790 (void *)&map->passive.flags)) { 791 req_id = READ_ONCE(map->passive.inflight_req_id); 792 if (req_id != PVCALLS_INVALID_ID && 793 READ_ONCE(bedata->rsp[req_id].req_id) == req_id) { 794 map2 = map->passive.accept_map; 795 goto received; 796 } 797 if (nonblock) { 798 pvcalls_exit_sock(sock); 799 return -EAGAIN; 800 } 801 if (wait_event_interruptible(map->passive.inflight_accept_req, 802 !test_and_set_bit(PVCALLS_FLAG_ACCEPT_INFLIGHT, 803 (void *)&map->passive.flags))) { 804 pvcalls_exit_sock(sock); 805 return -EINTR; 806 } 807 } 808 809 map2 = kzalloc(sizeof(*map2), GFP_KERNEL); 810 if (map2 == NULL) { 811 clear_bit(PVCALLS_FLAG_ACCEPT_INFLIGHT, 812 (void *)&map->passive.flags); 813 pvcalls_exit_sock(sock); 814 return -ENOMEM; 815 } 816 ret = alloc_active_ring(map2); 817 if (ret < 0) { 818 clear_bit(PVCALLS_FLAG_ACCEPT_INFLIGHT, 819 (void *)&map->passive.flags); 820 kfree(map2); 821 pvcalls_exit_sock(sock); 822 return ret; 823 } 824 spin_lock(&bedata->socket_lock); 825 ret = get_request(bedata, &req_id); 826 if (ret < 0) { 827 clear_bit(PVCALLS_FLAG_ACCEPT_INFLIGHT, 828 (void *)&map->passive.flags); 829 spin_unlock(&bedata->socket_lock); 830 free_active_ring(map2); 831 kfree(map2); 832 pvcalls_exit_sock(sock); 833 return ret; 834 } 835 836 ret = create_active(map2, &evtchn); 837 if (ret < 0) { 838 free_active_ring(map2); 839 kfree(map2); 840 clear_bit(PVCALLS_FLAG_ACCEPT_INFLIGHT, 841 (void *)&map->passive.flags); 842 spin_unlock(&bedata->socket_lock); 843 pvcalls_exit_sock(sock); 844 return ret; 845 } 846 list_add_tail(&map2->list, &bedata->socket_mappings); 847 848 req = RING_GET_REQUEST(&bedata->ring, req_id); 849 req->req_id = req_id; 850 req->cmd = PVCALLS_ACCEPT; 851 req->u.accept.id = (uintptr_t) map; 852 req->u.accept.ref = map2->active.ref; 853 req->u.accept.id_new = (uintptr_t) map2; 854 req->u.accept.evtchn = evtchn; 855 map->passive.accept_map = map2; 856 857 bedata->ring.req_prod_pvt++; 858 RING_PUSH_REQUESTS_AND_CHECK_NOTIFY(&bedata->ring, notify); 859 spin_unlock(&bedata->socket_lock); 860 if (notify) 861 notify_remote_via_irq(bedata->irq); 862 /* We could check if we have received a response before returning. */ 863 if (nonblock) { 864 WRITE_ONCE(map->passive.inflight_req_id, req_id); 865 pvcalls_exit_sock(sock); 866 return -EAGAIN; 867 } 868 869 if (wait_event_interruptible(bedata->inflight_req, 870 READ_ONCE(bedata->rsp[req_id].req_id) == req_id)) { 871 pvcalls_exit_sock(sock); 872 return -EINTR; 873 } 874 /* read req_id, then the content */ 875 smp_rmb(); 876 877 received: 878 map2->sock = newsock; 879 newsock->sk = sk_alloc(sock_net(sock->sk), PF_INET, GFP_KERNEL, &pvcalls_proto, false); 880 if (!newsock->sk) { 881 bedata->rsp[req_id].req_id = PVCALLS_INVALID_ID; 882 map->passive.inflight_req_id = PVCALLS_INVALID_ID; 883 clear_bit(PVCALLS_FLAG_ACCEPT_INFLIGHT, 884 (void *)&map->passive.flags); 885 pvcalls_front_free_map(bedata, map2); 886 pvcalls_exit_sock(sock); 887 return -ENOMEM; 888 } 889 newsock->sk->sk_send_head = (void *)map2; 890 891 ret = bedata->rsp[req_id].ret; 892 bedata->rsp[req_id].req_id = PVCALLS_INVALID_ID; 893 map->passive.inflight_req_id = PVCALLS_INVALID_ID; 894 895 clear_bit(PVCALLS_FLAG_ACCEPT_INFLIGHT, (void *)&map->passive.flags); 896 wake_up(&map->passive.inflight_accept_req); 897 898 pvcalls_exit_sock(sock); 899 return ret; 900 } 901 902 static __poll_t pvcalls_front_poll_passive(struct file *file, 903 struct pvcalls_bedata *bedata, 904 struct sock_mapping *map, 905 poll_table *wait) 906 { 907 int notify, req_id, ret; 908 struct xen_pvcalls_request *req; 909 910 if (test_bit(PVCALLS_FLAG_ACCEPT_INFLIGHT, 911 (void *)&map->passive.flags)) { 912 uint32_t req_id = READ_ONCE(map->passive.inflight_req_id); 913 914 if (req_id != PVCALLS_INVALID_ID && 915 READ_ONCE(bedata->rsp[req_id].req_id) == req_id) 916 return EPOLLIN | EPOLLRDNORM; 917 918 poll_wait(file, &map->passive.inflight_accept_req, wait); 919 return 0; 920 } 921 922 if (test_and_clear_bit(PVCALLS_FLAG_POLL_RET, 923 (void *)&map->passive.flags)) 924 return EPOLLIN | EPOLLRDNORM; 925 926 /* 927 * First check RET, then INFLIGHT. No barriers necessary to 928 * ensure execution ordering because of the conditional 929 * instructions creating control dependencies. 930 */ 931 932 if (test_and_set_bit(PVCALLS_FLAG_POLL_INFLIGHT, 933 (void *)&map->passive.flags)) { 934 poll_wait(file, &bedata->inflight_req, wait); 935 return 0; 936 } 937 938 spin_lock(&bedata->socket_lock); 939 ret = get_request(bedata, &req_id); 940 if (ret < 0) { 941 spin_unlock(&bedata->socket_lock); 942 return ret; 943 } 944 req = RING_GET_REQUEST(&bedata->ring, req_id); 945 req->req_id = req_id; 946 req->cmd = PVCALLS_POLL; 947 req->u.poll.id = (uintptr_t) map; 948 949 bedata->ring.req_prod_pvt++; 950 RING_PUSH_REQUESTS_AND_CHECK_NOTIFY(&bedata->ring, notify); 951 spin_unlock(&bedata->socket_lock); 952 if (notify) 953 notify_remote_via_irq(bedata->irq); 954 955 poll_wait(file, &bedata->inflight_req, wait); 956 return 0; 957 } 958 959 static __poll_t pvcalls_front_poll_active(struct file *file, 960 struct pvcalls_bedata *bedata, 961 struct sock_mapping *map, 962 poll_table *wait) 963 { 964 __poll_t mask = 0; 965 int32_t in_error, out_error; 966 struct pvcalls_data_intf *intf = map->active.ring; 967 968 out_error = intf->out_error; 969 in_error = intf->in_error; 970 971 poll_wait(file, &map->active.inflight_conn_req, wait); 972 if (pvcalls_front_write_todo(map)) 973 mask |= EPOLLOUT | EPOLLWRNORM; 974 if (pvcalls_front_read_todo(map)) 975 mask |= EPOLLIN | EPOLLRDNORM; 976 if (in_error != 0 || out_error != 0) 977 mask |= EPOLLERR; 978 979 return mask; 980 } 981 982 __poll_t pvcalls_front_poll(struct file *file, struct socket *sock, 983 poll_table *wait) 984 { 985 struct pvcalls_bedata *bedata; 986 struct sock_mapping *map; 987 __poll_t ret; 988 989 map = pvcalls_enter_sock(sock); 990 if (IS_ERR(map)) 991 return EPOLLNVAL; 992 bedata = dev_get_drvdata(&pvcalls_front_dev->dev); 993 994 if (map->active_socket) 995 ret = pvcalls_front_poll_active(file, bedata, map, wait); 996 else 997 ret = pvcalls_front_poll_passive(file, bedata, map, wait); 998 pvcalls_exit_sock(sock); 999 return ret; 1000 } 1001 1002 int pvcalls_front_release(struct socket *sock) 1003 { 1004 struct pvcalls_bedata *bedata; 1005 struct sock_mapping *map; 1006 int req_id, notify, ret; 1007 struct xen_pvcalls_request *req; 1008 1009 if (sock->sk == NULL) 1010 return 0; 1011 1012 map = pvcalls_enter_sock(sock); 1013 if (IS_ERR(map)) { 1014 if (PTR_ERR(map) == -ENOTCONN) 1015 return -EIO; 1016 else 1017 return 0; 1018 } 1019 bedata = dev_get_drvdata(&pvcalls_front_dev->dev); 1020 1021 spin_lock(&bedata->socket_lock); 1022 ret = get_request(bedata, &req_id); 1023 if (ret < 0) { 1024 spin_unlock(&bedata->socket_lock); 1025 pvcalls_exit_sock(sock); 1026 return ret; 1027 } 1028 sock->sk->sk_send_head = NULL; 1029 1030 req = RING_GET_REQUEST(&bedata->ring, req_id); 1031 req->req_id = req_id; 1032 req->cmd = PVCALLS_RELEASE; 1033 req->u.release.id = (uintptr_t)map; 1034 1035 bedata->ring.req_prod_pvt++; 1036 RING_PUSH_REQUESTS_AND_CHECK_NOTIFY(&bedata->ring, notify); 1037 spin_unlock(&bedata->socket_lock); 1038 if (notify) 1039 notify_remote_via_irq(bedata->irq); 1040 1041 wait_event(bedata->inflight_req, 1042 READ_ONCE(bedata->rsp[req_id].req_id) == req_id); 1043 1044 if (map->active_socket) { 1045 /* 1046 * Set in_error and wake up inflight_conn_req to force 1047 * recvmsg waiters to exit. 1048 */ 1049 map->active.ring->in_error = -EBADF; 1050 wake_up_interruptible(&map->active.inflight_conn_req); 1051 1052 /* 1053 * We need to make sure that sendmsg/recvmsg on this socket have 1054 * not started before we've cleared sk_send_head here. The 1055 * easiest way to guarantee this is to see that no pvcalls 1056 * (other than us) is in progress on this socket. 1057 */ 1058 while (atomic_read(&map->refcount) > 1) 1059 cpu_relax(); 1060 1061 pvcalls_front_free_map(bedata, map); 1062 } else { 1063 wake_up(&bedata->inflight_req); 1064 wake_up(&map->passive.inflight_accept_req); 1065 1066 while (atomic_read(&map->refcount) > 1) 1067 cpu_relax(); 1068 1069 spin_lock(&bedata->socket_lock); 1070 list_del(&map->list); 1071 spin_unlock(&bedata->socket_lock); 1072 if (READ_ONCE(map->passive.inflight_req_id) != PVCALLS_INVALID_ID && 1073 READ_ONCE(map->passive.inflight_req_id) != 0) { 1074 pvcalls_front_free_map(bedata, 1075 map->passive.accept_map); 1076 } 1077 kfree(map); 1078 } 1079 WRITE_ONCE(bedata->rsp[req_id].req_id, PVCALLS_INVALID_ID); 1080 1081 pvcalls_exit(); 1082 return 0; 1083 } 1084 1085 static const struct xenbus_device_id pvcalls_front_ids[] = { 1086 { "pvcalls" }, 1087 { "" } 1088 }; 1089 1090 static int pvcalls_front_remove(struct xenbus_device *dev) 1091 { 1092 struct pvcalls_bedata *bedata; 1093 struct sock_mapping *map = NULL, *n; 1094 1095 bedata = dev_get_drvdata(&pvcalls_front_dev->dev); 1096 dev_set_drvdata(&dev->dev, NULL); 1097 pvcalls_front_dev = NULL; 1098 if (bedata->irq >= 0) 1099 unbind_from_irqhandler(bedata->irq, dev); 1100 1101 list_for_each_entry_safe(map, n, &bedata->socket_mappings, list) { 1102 map->sock->sk->sk_send_head = NULL; 1103 if (map->active_socket) { 1104 map->active.ring->in_error = -EBADF; 1105 wake_up_interruptible(&map->active.inflight_conn_req); 1106 } 1107 } 1108 1109 smp_mb(); 1110 while (atomic_read(&pvcalls_refcount) > 0) 1111 cpu_relax(); 1112 list_for_each_entry_safe(map, n, &bedata->socket_mappings, list) { 1113 if (map->active_socket) { 1114 /* No need to lock, refcount is 0 */ 1115 pvcalls_front_free_map(bedata, map); 1116 } else { 1117 list_del(&map->list); 1118 kfree(map); 1119 } 1120 } 1121 if (bedata->ref != -1) 1122 gnttab_end_foreign_access(bedata->ref, 0, 0); 1123 kfree(bedata->ring.sring); 1124 kfree(bedata); 1125 xenbus_switch_state(dev, XenbusStateClosed); 1126 return 0; 1127 } 1128 1129 static int pvcalls_front_probe(struct xenbus_device *dev, 1130 const struct xenbus_device_id *id) 1131 { 1132 int ret = -ENOMEM, evtchn, i; 1133 unsigned int max_page_order, function_calls, len; 1134 char *versions; 1135 grant_ref_t gref_head = 0; 1136 struct xenbus_transaction xbt; 1137 struct pvcalls_bedata *bedata = NULL; 1138 struct xen_pvcalls_sring *sring; 1139 1140 if (pvcalls_front_dev != NULL) { 1141 dev_err(&dev->dev, "only one PV Calls connection supported\n"); 1142 return -EINVAL; 1143 } 1144 1145 versions = xenbus_read(XBT_NIL, dev->otherend, "versions", &len); 1146 if (IS_ERR(versions)) 1147 return PTR_ERR(versions); 1148 if (!len) 1149 return -EINVAL; 1150 if (strcmp(versions, "1")) { 1151 kfree(versions); 1152 return -EINVAL; 1153 } 1154 kfree(versions); 1155 max_page_order = xenbus_read_unsigned(dev->otherend, 1156 "max-page-order", 0); 1157 if (max_page_order < PVCALLS_RING_ORDER) 1158 return -ENODEV; 1159 function_calls = xenbus_read_unsigned(dev->otherend, 1160 "function-calls", 0); 1161 /* See XENBUS_FUNCTIONS_CALLS in pvcalls.h */ 1162 if (function_calls != 1) 1163 return -ENODEV; 1164 pr_info("%s max-page-order is %u\n", __func__, max_page_order); 1165 1166 bedata = kzalloc(sizeof(struct pvcalls_bedata), GFP_KERNEL); 1167 if (!bedata) 1168 return -ENOMEM; 1169 1170 dev_set_drvdata(&dev->dev, bedata); 1171 pvcalls_front_dev = dev; 1172 init_waitqueue_head(&bedata->inflight_req); 1173 INIT_LIST_HEAD(&bedata->socket_mappings); 1174 spin_lock_init(&bedata->socket_lock); 1175 bedata->irq = -1; 1176 bedata->ref = -1; 1177 1178 for (i = 0; i < PVCALLS_NR_RSP_PER_RING; i++) 1179 bedata->rsp[i].req_id = PVCALLS_INVALID_ID; 1180 1181 sring = (struct xen_pvcalls_sring *) __get_free_page(GFP_KERNEL | 1182 __GFP_ZERO); 1183 if (!sring) 1184 goto error; 1185 SHARED_RING_INIT(sring); 1186 FRONT_RING_INIT(&bedata->ring, sring, XEN_PAGE_SIZE); 1187 1188 ret = xenbus_alloc_evtchn(dev, &evtchn); 1189 if (ret) 1190 goto error; 1191 1192 bedata->irq = bind_evtchn_to_irqhandler(evtchn, 1193 pvcalls_front_event_handler, 1194 0, "pvcalls-frontend", dev); 1195 if (bedata->irq < 0) { 1196 ret = bedata->irq; 1197 goto error; 1198 } 1199 1200 ret = gnttab_alloc_grant_references(1, &gref_head); 1201 if (ret < 0) 1202 goto error; 1203 ret = gnttab_claim_grant_reference(&gref_head); 1204 if (ret < 0) 1205 goto error; 1206 bedata->ref = ret; 1207 gnttab_grant_foreign_access_ref(bedata->ref, dev->otherend_id, 1208 virt_to_gfn((void *)sring), 0); 1209 1210 again: 1211 ret = xenbus_transaction_start(&xbt); 1212 if (ret) { 1213 xenbus_dev_fatal(dev, ret, "starting transaction"); 1214 goto error; 1215 } 1216 ret = xenbus_printf(xbt, dev->nodename, "version", "%u", 1); 1217 if (ret) 1218 goto error_xenbus; 1219 ret = xenbus_printf(xbt, dev->nodename, "ring-ref", "%d", bedata->ref); 1220 if (ret) 1221 goto error_xenbus; 1222 ret = xenbus_printf(xbt, dev->nodename, "port", "%u", 1223 evtchn); 1224 if (ret) 1225 goto error_xenbus; 1226 ret = xenbus_transaction_end(xbt, 0); 1227 if (ret) { 1228 if (ret == -EAGAIN) 1229 goto again; 1230 xenbus_dev_fatal(dev, ret, "completing transaction"); 1231 goto error; 1232 } 1233 xenbus_switch_state(dev, XenbusStateInitialised); 1234 1235 return 0; 1236 1237 error_xenbus: 1238 xenbus_transaction_end(xbt, 1); 1239 xenbus_dev_fatal(dev, ret, "writing xenstore"); 1240 error: 1241 pvcalls_front_remove(dev); 1242 return ret; 1243 } 1244 1245 static void pvcalls_front_changed(struct xenbus_device *dev, 1246 enum xenbus_state backend_state) 1247 { 1248 switch (backend_state) { 1249 case XenbusStateReconfiguring: 1250 case XenbusStateReconfigured: 1251 case XenbusStateInitialising: 1252 case XenbusStateInitialised: 1253 case XenbusStateUnknown: 1254 break; 1255 1256 case XenbusStateInitWait: 1257 break; 1258 1259 case XenbusStateConnected: 1260 xenbus_switch_state(dev, XenbusStateConnected); 1261 break; 1262 1263 case XenbusStateClosed: 1264 if (dev->state == XenbusStateClosed) 1265 break; 1266 /* Missed the backend's CLOSING state */ 1267 /* fall through */ 1268 case XenbusStateClosing: 1269 xenbus_frontend_closed(dev); 1270 break; 1271 } 1272 } 1273 1274 static struct xenbus_driver pvcalls_front_driver = { 1275 .ids = pvcalls_front_ids, 1276 .probe = pvcalls_front_probe, 1277 .remove = pvcalls_front_remove, 1278 .otherend_changed = pvcalls_front_changed, 1279 }; 1280 1281 static int __init pvcalls_frontend_init(void) 1282 { 1283 if (!xen_domain()) 1284 return -ENODEV; 1285 1286 pr_info("Initialising Xen pvcalls frontend driver\n"); 1287 1288 return xenbus_register_frontend(&pvcalls_front_driver); 1289 } 1290 1291 module_init(pvcalls_frontend_init); 1292 1293 MODULE_DESCRIPTION("Xen PV Calls frontend driver"); 1294 MODULE_AUTHOR("Stefano Stabellini <sstabellini@kernel.org>"); 1295 MODULE_LICENSE("GPL"); 1296