1 /* 2 * Copyright (c) 2005 Topspin Communications. All rights reserved. 3 * Copyright (c) 2005, 2006 Cisco Systems. All rights reserved. 4 * Copyright (c) 2005 Mellanox Technologies. All rights reserved. 5 * Copyright (c) 2005 Voltaire, Inc. All rights reserved. 6 * Copyright (c) 2005 PathScale, Inc. All rights reserved. 7 * 8 * This software is available to you under a choice of one of two 9 * licenses. You may choose to be licensed under the terms of the GNU 10 * General Public License (GPL) Version 2, available from the file 11 * COPYING in the main directory of this source tree, or the 12 * OpenIB.org BSD license below: 13 * 14 * Redistribution and use in source and binary forms, with or 15 * without modification, are permitted provided that the following 16 * conditions are met: 17 * 18 * - Redistributions of source code must retain the above 19 * copyright notice, this list of conditions and the following 20 * disclaimer. 21 * 22 * - Redistributions in binary form must reproduce the above 23 * copyright notice, this list of conditions and the following 24 * disclaimer in the documentation and/or other materials 25 * provided with the distribution. 26 * 27 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, 28 * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF 29 * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND 30 * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS 31 * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN 32 * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN 33 * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE 34 * SOFTWARE. 35 */ 36 37 #include <linux/module.h> 38 #include <linux/init.h> 39 #include <linux/device.h> 40 #include <linux/err.h> 41 #include <linux/fs.h> 42 #include <linux/poll.h> 43 #include <linux/sched.h> 44 #include <linux/file.h> 45 #include <linux/cdev.h> 46 #include <linux/anon_inodes.h> 47 #include <linux/slab.h> 48 #include <linux/sched/mm.h> 49 50 #include <linux/uaccess.h> 51 52 #include <rdma/ib.h> 53 #include <rdma/uverbs_std_types.h> 54 #include <rdma/rdma_netlink.h> 55 56 #include "uverbs.h" 57 #include "core_priv.h" 58 #include "rdma_core.h" 59 60 MODULE_AUTHOR("Roland Dreier"); 61 MODULE_DESCRIPTION("InfiniBand userspace verbs access"); 62 MODULE_LICENSE("Dual BSD/GPL"); 63 64 enum { 65 IB_UVERBS_MAJOR = 231, 66 IB_UVERBS_BASE_MINOR = 192, 67 IB_UVERBS_MAX_DEVICES = RDMA_MAX_PORTS, 68 IB_UVERBS_NUM_FIXED_MINOR = 32, 69 IB_UVERBS_NUM_DYNAMIC_MINOR = IB_UVERBS_MAX_DEVICES - IB_UVERBS_NUM_FIXED_MINOR, 70 }; 71 72 #define IB_UVERBS_BASE_DEV MKDEV(IB_UVERBS_MAJOR, IB_UVERBS_BASE_MINOR) 73 74 static dev_t dynamic_uverbs_dev; 75 static struct class *uverbs_class; 76 77 static DEFINE_IDA(uverbs_ida); 78 static void ib_uverbs_add_one(struct ib_device *device); 79 static void ib_uverbs_remove_one(struct ib_device *device, void *client_data); 80 81 /* 82 * Must be called with the ufile->device->disassociate_srcu held, and the lock 83 * must be held until use of the ucontext is finished. 84 */ 85 struct ib_ucontext *ib_uverbs_get_ucontext_file(struct ib_uverbs_file *ufile) 86 { 87 /* 88 * We do not hold the hw_destroy_rwsem lock for this flow, instead 89 * srcu is used. It does not matter if someone races this with 90 * get_context, we get NULL or valid ucontext. 91 */ 92 struct ib_ucontext *ucontext = smp_load_acquire(&ufile->ucontext); 93 94 if (!srcu_dereference(ufile->device->ib_dev, 95 &ufile->device->disassociate_srcu)) 96 return ERR_PTR(-EIO); 97 98 if (!ucontext) 99 return ERR_PTR(-EINVAL); 100 101 return ucontext; 102 } 103 EXPORT_SYMBOL(ib_uverbs_get_ucontext_file); 104 105 int uverbs_dealloc_mw(struct ib_mw *mw) 106 { 107 struct ib_pd *pd = mw->pd; 108 int ret; 109 110 ret = mw->device->ops.dealloc_mw(mw); 111 if (!ret) 112 atomic_dec(&pd->usecnt); 113 return ret; 114 } 115 116 static void ib_uverbs_release_dev(struct device *device) 117 { 118 struct ib_uverbs_device *dev = 119 container_of(device, struct ib_uverbs_device, dev); 120 121 uverbs_destroy_api(dev->uapi); 122 cleanup_srcu_struct(&dev->disassociate_srcu); 123 mutex_destroy(&dev->lists_mutex); 124 mutex_destroy(&dev->xrcd_tree_mutex); 125 kfree(dev); 126 } 127 128 void ib_uverbs_release_ucq(struct ib_uverbs_completion_event_file *ev_file, 129 struct ib_ucq_object *uobj) 130 { 131 struct ib_uverbs_event *evt, *tmp; 132 133 if (ev_file) { 134 spin_lock_irq(&ev_file->ev_queue.lock); 135 list_for_each_entry_safe(evt, tmp, &uobj->comp_list, obj_list) { 136 list_del(&evt->list); 137 kfree(evt); 138 } 139 spin_unlock_irq(&ev_file->ev_queue.lock); 140 141 uverbs_uobject_put(&ev_file->uobj); 142 } 143 144 ib_uverbs_release_uevent(&uobj->uevent); 145 } 146 147 void ib_uverbs_release_uevent(struct ib_uevent_object *uobj) 148 { 149 struct ib_uverbs_async_event_file *async_file = 150 READ_ONCE(uobj->uobject.ufile->async_file); 151 struct ib_uverbs_event *evt, *tmp; 152 153 if (!async_file) 154 return; 155 156 spin_lock_irq(&async_file->ev_queue.lock); 157 list_for_each_entry_safe(evt, tmp, &uobj->event_list, obj_list) { 158 list_del(&evt->list); 159 kfree(evt); 160 } 161 spin_unlock_irq(&async_file->ev_queue.lock); 162 } 163 164 void ib_uverbs_detach_umcast(struct ib_qp *qp, 165 struct ib_uqp_object *uobj) 166 { 167 struct ib_uverbs_mcast_entry *mcast, *tmp; 168 169 list_for_each_entry_safe(mcast, tmp, &uobj->mcast_list, list) { 170 ib_detach_mcast(qp, &mcast->gid, mcast->lid); 171 list_del(&mcast->list); 172 kfree(mcast); 173 } 174 } 175 176 static void ib_uverbs_comp_dev(struct ib_uverbs_device *dev) 177 { 178 complete(&dev->comp); 179 } 180 181 void ib_uverbs_release_file(struct kref *ref) 182 { 183 struct ib_uverbs_file *file = 184 container_of(ref, struct ib_uverbs_file, ref); 185 struct ib_device *ib_dev; 186 int srcu_key; 187 188 release_ufile_idr_uobject(file); 189 190 srcu_key = srcu_read_lock(&file->device->disassociate_srcu); 191 ib_dev = srcu_dereference(file->device->ib_dev, 192 &file->device->disassociate_srcu); 193 if (ib_dev && !ib_dev->ops.disassociate_ucontext) 194 module_put(ib_dev->ops.owner); 195 srcu_read_unlock(&file->device->disassociate_srcu, srcu_key); 196 197 if (atomic_dec_and_test(&file->device->refcount)) 198 ib_uverbs_comp_dev(file->device); 199 200 if (file->async_file) 201 uverbs_uobject_put(&file->async_file->uobj); 202 put_device(&file->device->dev); 203 204 if (file->disassociate_page) 205 __free_pages(file->disassociate_page, 0); 206 mutex_destroy(&file->umap_lock); 207 mutex_destroy(&file->ucontext_lock); 208 kfree(file); 209 } 210 211 static ssize_t ib_uverbs_event_read(struct ib_uverbs_event_queue *ev_queue, 212 struct file *filp, char __user *buf, 213 size_t count, loff_t *pos, 214 size_t eventsz) 215 { 216 struct ib_uverbs_event *event; 217 int ret = 0; 218 219 spin_lock_irq(&ev_queue->lock); 220 221 while (list_empty(&ev_queue->event_list)) { 222 spin_unlock_irq(&ev_queue->lock); 223 224 if (filp->f_flags & O_NONBLOCK) 225 return -EAGAIN; 226 227 if (wait_event_interruptible(ev_queue->poll_wait, 228 (!list_empty(&ev_queue->event_list) || 229 ev_queue->is_closed))) 230 return -ERESTARTSYS; 231 232 spin_lock_irq(&ev_queue->lock); 233 234 /* If device was disassociated and no event exists set an error */ 235 if (list_empty(&ev_queue->event_list) && ev_queue->is_closed) { 236 spin_unlock_irq(&ev_queue->lock); 237 return -EIO; 238 } 239 } 240 241 event = list_entry(ev_queue->event_list.next, struct ib_uverbs_event, list); 242 243 if (eventsz > count) { 244 ret = -EINVAL; 245 event = NULL; 246 } else { 247 list_del(ev_queue->event_list.next); 248 if (event->counter) { 249 ++(*event->counter); 250 list_del(&event->obj_list); 251 } 252 } 253 254 spin_unlock_irq(&ev_queue->lock); 255 256 if (event) { 257 if (copy_to_user(buf, event, eventsz)) 258 ret = -EFAULT; 259 else 260 ret = eventsz; 261 } 262 263 kfree(event); 264 265 return ret; 266 } 267 268 static ssize_t ib_uverbs_async_event_read(struct file *filp, char __user *buf, 269 size_t count, loff_t *pos) 270 { 271 struct ib_uverbs_async_event_file *file = filp->private_data; 272 273 return ib_uverbs_event_read(&file->ev_queue, filp, buf, count, pos, 274 sizeof(struct ib_uverbs_async_event_desc)); 275 } 276 277 static ssize_t ib_uverbs_comp_event_read(struct file *filp, char __user *buf, 278 size_t count, loff_t *pos) 279 { 280 struct ib_uverbs_completion_event_file *comp_ev_file = 281 filp->private_data; 282 283 return ib_uverbs_event_read(&comp_ev_file->ev_queue, filp, buf, count, 284 pos, 285 sizeof(struct ib_uverbs_comp_event_desc)); 286 } 287 288 static __poll_t ib_uverbs_event_poll(struct ib_uverbs_event_queue *ev_queue, 289 struct file *filp, 290 struct poll_table_struct *wait) 291 { 292 __poll_t pollflags = 0; 293 294 poll_wait(filp, &ev_queue->poll_wait, wait); 295 296 spin_lock_irq(&ev_queue->lock); 297 if (!list_empty(&ev_queue->event_list)) 298 pollflags = EPOLLIN | EPOLLRDNORM; 299 spin_unlock_irq(&ev_queue->lock); 300 301 return pollflags; 302 } 303 304 static __poll_t ib_uverbs_async_event_poll(struct file *filp, 305 struct poll_table_struct *wait) 306 { 307 struct ib_uverbs_async_event_file *file = filp->private_data; 308 309 return ib_uverbs_event_poll(&file->ev_queue, filp, wait); 310 } 311 312 static __poll_t ib_uverbs_comp_event_poll(struct file *filp, 313 struct poll_table_struct *wait) 314 { 315 struct ib_uverbs_completion_event_file *comp_ev_file = 316 filp->private_data; 317 318 return ib_uverbs_event_poll(&comp_ev_file->ev_queue, filp, wait); 319 } 320 321 static int ib_uverbs_async_event_fasync(int fd, struct file *filp, int on) 322 { 323 struct ib_uverbs_async_event_file *file = filp->private_data; 324 325 return fasync_helper(fd, filp, on, &file->ev_queue.async_queue); 326 } 327 328 static int ib_uverbs_comp_event_fasync(int fd, struct file *filp, int on) 329 { 330 struct ib_uverbs_completion_event_file *comp_ev_file = 331 filp->private_data; 332 333 return fasync_helper(fd, filp, on, &comp_ev_file->ev_queue.async_queue); 334 } 335 336 const struct file_operations uverbs_event_fops = { 337 .owner = THIS_MODULE, 338 .read = ib_uverbs_comp_event_read, 339 .poll = ib_uverbs_comp_event_poll, 340 .release = uverbs_uobject_fd_release, 341 .fasync = ib_uverbs_comp_event_fasync, 342 .llseek = no_llseek, 343 }; 344 345 const struct file_operations uverbs_async_event_fops = { 346 .owner = THIS_MODULE, 347 .read = ib_uverbs_async_event_read, 348 .poll = ib_uverbs_async_event_poll, 349 .release = uverbs_uobject_fd_release, 350 .fasync = ib_uverbs_async_event_fasync, 351 .llseek = no_llseek, 352 }; 353 354 void ib_uverbs_comp_handler(struct ib_cq *cq, void *cq_context) 355 { 356 struct ib_uverbs_event_queue *ev_queue = cq_context; 357 struct ib_ucq_object *uobj; 358 struct ib_uverbs_event *entry; 359 unsigned long flags; 360 361 if (!ev_queue) 362 return; 363 364 spin_lock_irqsave(&ev_queue->lock, flags); 365 if (ev_queue->is_closed) { 366 spin_unlock_irqrestore(&ev_queue->lock, flags); 367 return; 368 } 369 370 entry = kmalloc(sizeof(*entry), GFP_ATOMIC); 371 if (!entry) { 372 spin_unlock_irqrestore(&ev_queue->lock, flags); 373 return; 374 } 375 376 uobj = cq->uobject; 377 378 entry->desc.comp.cq_handle = cq->uobject->uevent.uobject.user_handle; 379 entry->counter = &uobj->comp_events_reported; 380 381 list_add_tail(&entry->list, &ev_queue->event_list); 382 list_add_tail(&entry->obj_list, &uobj->comp_list); 383 spin_unlock_irqrestore(&ev_queue->lock, flags); 384 385 wake_up_interruptible(&ev_queue->poll_wait); 386 kill_fasync(&ev_queue->async_queue, SIGIO, POLL_IN); 387 } 388 389 static void 390 ib_uverbs_async_handler(struct ib_uverbs_async_event_file *async_file, 391 __u64 element, __u64 event, struct list_head *obj_list, 392 u32 *counter) 393 { 394 struct ib_uverbs_event *entry; 395 unsigned long flags; 396 397 if (!async_file) 398 return; 399 400 spin_lock_irqsave(&async_file->ev_queue.lock, flags); 401 if (async_file->ev_queue.is_closed) { 402 spin_unlock_irqrestore(&async_file->ev_queue.lock, flags); 403 return; 404 } 405 406 entry = kmalloc(sizeof(*entry), GFP_ATOMIC); 407 if (!entry) { 408 spin_unlock_irqrestore(&async_file->ev_queue.lock, flags); 409 return; 410 } 411 412 entry->desc.async.element = element; 413 entry->desc.async.event_type = event; 414 entry->desc.async.reserved = 0; 415 entry->counter = counter; 416 417 list_add_tail(&entry->list, &async_file->ev_queue.event_list); 418 if (obj_list) 419 list_add_tail(&entry->obj_list, obj_list); 420 spin_unlock_irqrestore(&async_file->ev_queue.lock, flags); 421 422 wake_up_interruptible(&async_file->ev_queue.poll_wait); 423 kill_fasync(&async_file->ev_queue.async_queue, SIGIO, POLL_IN); 424 } 425 426 static void uverbs_uobj_event(struct ib_uevent_object *eobj, 427 struct ib_event *event) 428 { 429 ib_uverbs_async_handler(READ_ONCE(eobj->uobject.ufile->async_file), 430 eobj->uobject.user_handle, event->event, 431 &eobj->event_list, &eobj->events_reported); 432 } 433 434 void ib_uverbs_cq_event_handler(struct ib_event *event, void *context_ptr) 435 { 436 uverbs_uobj_event(&event->element.cq->uobject->uevent, event); 437 } 438 439 void ib_uverbs_qp_event_handler(struct ib_event *event, void *context_ptr) 440 { 441 /* for XRC target qp's, check that qp is live */ 442 if (!event->element.qp->uobject) 443 return; 444 445 uverbs_uobj_event(&event->element.qp->uobject->uevent, event); 446 } 447 448 void ib_uverbs_wq_event_handler(struct ib_event *event, void *context_ptr) 449 { 450 uverbs_uobj_event(&event->element.wq->uobject->uevent, event); 451 } 452 453 void ib_uverbs_srq_event_handler(struct ib_event *event, void *context_ptr) 454 { 455 uverbs_uobj_event(&event->element.srq->uobject->uevent, event); 456 } 457 458 static void ib_uverbs_event_handler(struct ib_event_handler *handler, 459 struct ib_event *event) 460 { 461 ib_uverbs_async_handler( 462 container_of(handler, struct ib_uverbs_async_event_file, 463 event_handler), 464 event->element.port_num, event->event, NULL, NULL); 465 } 466 467 void ib_uverbs_init_event_queue(struct ib_uverbs_event_queue *ev_queue) 468 { 469 spin_lock_init(&ev_queue->lock); 470 INIT_LIST_HEAD(&ev_queue->event_list); 471 init_waitqueue_head(&ev_queue->poll_wait); 472 ev_queue->is_closed = 0; 473 ev_queue->async_queue = NULL; 474 } 475 476 void ib_uverbs_init_async_event_file( 477 struct ib_uverbs_async_event_file *async_file) 478 { 479 struct ib_uverbs_file *uverbs_file = async_file->uobj.ufile; 480 struct ib_device *ib_dev = async_file->uobj.context->device; 481 482 ib_uverbs_init_event_queue(&async_file->ev_queue); 483 484 /* The first async_event_file becomes the default one for the file. */ 485 mutex_lock(&uverbs_file->ucontext_lock); 486 if (!uverbs_file->async_file) { 487 /* Pairs with the put in ib_uverbs_release_file */ 488 uverbs_uobject_get(&async_file->uobj); 489 smp_store_release(&uverbs_file->async_file, async_file); 490 } 491 mutex_unlock(&uverbs_file->ucontext_lock); 492 493 INIT_IB_EVENT_HANDLER(&async_file->event_handler, ib_dev, 494 ib_uverbs_event_handler); 495 ib_register_event_handler(&async_file->event_handler); 496 } 497 498 static ssize_t verify_hdr(struct ib_uverbs_cmd_hdr *hdr, 499 struct ib_uverbs_ex_cmd_hdr *ex_hdr, size_t count, 500 const struct uverbs_api_write_method *method_elm) 501 { 502 if (method_elm->is_ex) { 503 count -= sizeof(*hdr) + sizeof(*ex_hdr); 504 505 if ((hdr->in_words + ex_hdr->provider_in_words) * 8 != count) 506 return -EINVAL; 507 508 if (hdr->in_words * 8 < method_elm->req_size) 509 return -ENOSPC; 510 511 if (ex_hdr->cmd_hdr_reserved) 512 return -EINVAL; 513 514 if (ex_hdr->response) { 515 if (!hdr->out_words && !ex_hdr->provider_out_words) 516 return -EINVAL; 517 518 if (hdr->out_words * 8 < method_elm->resp_size) 519 return -ENOSPC; 520 521 if (!access_ok(u64_to_user_ptr(ex_hdr->response), 522 (hdr->out_words + ex_hdr->provider_out_words) * 8)) 523 return -EFAULT; 524 } else { 525 if (hdr->out_words || ex_hdr->provider_out_words) 526 return -EINVAL; 527 } 528 529 return 0; 530 } 531 532 /* not extended command */ 533 if (hdr->in_words * 4 != count) 534 return -EINVAL; 535 536 if (count < method_elm->req_size + sizeof(hdr)) { 537 /* 538 * rdma-core v18 and v19 have a bug where they send DESTROY_CQ 539 * with a 16 byte write instead of 24. Old kernels didn't 540 * check the size so they allowed this. Now that the size is 541 * checked provide a compatibility work around to not break 542 * those userspaces. 543 */ 544 if (hdr->command == IB_USER_VERBS_CMD_DESTROY_CQ && 545 count == 16) { 546 hdr->in_words = 6; 547 return 0; 548 } 549 return -ENOSPC; 550 } 551 if (hdr->out_words * 4 < method_elm->resp_size) 552 return -ENOSPC; 553 554 return 0; 555 } 556 557 static ssize_t ib_uverbs_write(struct file *filp, const char __user *buf, 558 size_t count, loff_t *pos) 559 { 560 struct ib_uverbs_file *file = filp->private_data; 561 const struct uverbs_api_write_method *method_elm; 562 struct uverbs_api *uapi = file->device->uapi; 563 struct ib_uverbs_ex_cmd_hdr ex_hdr; 564 struct ib_uverbs_cmd_hdr hdr; 565 struct uverbs_attr_bundle bundle; 566 int srcu_key; 567 ssize_t ret; 568 569 if (!ib_safe_file_access(filp)) { 570 pr_err_once("uverbs_write: process %d (%s) changed security contexts after opening file descriptor, this is not allowed.\n", 571 task_tgid_vnr(current), current->comm); 572 return -EACCES; 573 } 574 575 if (count < sizeof(hdr)) 576 return -EINVAL; 577 578 if (copy_from_user(&hdr, buf, sizeof(hdr))) 579 return -EFAULT; 580 581 method_elm = uapi_get_method(uapi, hdr.command); 582 if (IS_ERR(method_elm)) 583 return PTR_ERR(method_elm); 584 585 if (method_elm->is_ex) { 586 if (count < (sizeof(hdr) + sizeof(ex_hdr))) 587 return -EINVAL; 588 if (copy_from_user(&ex_hdr, buf + sizeof(hdr), sizeof(ex_hdr))) 589 return -EFAULT; 590 } 591 592 ret = verify_hdr(&hdr, &ex_hdr, count, method_elm); 593 if (ret) 594 return ret; 595 596 srcu_key = srcu_read_lock(&file->device->disassociate_srcu); 597 598 buf += sizeof(hdr); 599 600 memset(bundle.attr_present, 0, sizeof(bundle.attr_present)); 601 bundle.ufile = file; 602 bundle.context = NULL; /* only valid if bundle has uobject */ 603 if (!method_elm->is_ex) { 604 size_t in_len = hdr.in_words * 4 - sizeof(hdr); 605 size_t out_len = hdr.out_words * 4; 606 u64 response = 0; 607 608 if (method_elm->has_udata) { 609 bundle.driver_udata.inlen = 610 in_len - method_elm->req_size; 611 in_len = method_elm->req_size; 612 if (bundle.driver_udata.inlen) 613 bundle.driver_udata.inbuf = buf + in_len; 614 else 615 bundle.driver_udata.inbuf = NULL; 616 } else { 617 memset(&bundle.driver_udata, 0, 618 sizeof(bundle.driver_udata)); 619 } 620 621 if (method_elm->has_resp) { 622 /* 623 * The macros check that if has_resp is set 624 * then the command request structure starts 625 * with a '__aligned u64 response' member. 626 */ 627 ret = get_user(response, (const u64 __user *)buf); 628 if (ret) 629 goto out_unlock; 630 631 if (method_elm->has_udata) { 632 bundle.driver_udata.outlen = 633 out_len - method_elm->resp_size; 634 out_len = method_elm->resp_size; 635 if (bundle.driver_udata.outlen) 636 bundle.driver_udata.outbuf = 637 u64_to_user_ptr(response + 638 out_len); 639 else 640 bundle.driver_udata.outbuf = NULL; 641 } 642 } else { 643 bundle.driver_udata.outlen = 0; 644 bundle.driver_udata.outbuf = NULL; 645 } 646 647 ib_uverbs_init_udata_buf_or_null( 648 &bundle.ucore, buf, u64_to_user_ptr(response), 649 in_len, out_len); 650 } else { 651 buf += sizeof(ex_hdr); 652 653 ib_uverbs_init_udata_buf_or_null(&bundle.ucore, buf, 654 u64_to_user_ptr(ex_hdr.response), 655 hdr.in_words * 8, hdr.out_words * 8); 656 657 ib_uverbs_init_udata_buf_or_null( 658 &bundle.driver_udata, buf + bundle.ucore.inlen, 659 u64_to_user_ptr(ex_hdr.response) + bundle.ucore.outlen, 660 ex_hdr.provider_in_words * 8, 661 ex_hdr.provider_out_words * 8); 662 663 } 664 665 ret = method_elm->handler(&bundle); 666 out_unlock: 667 srcu_read_unlock(&file->device->disassociate_srcu, srcu_key); 668 return (ret) ? : count; 669 } 670 671 static const struct vm_operations_struct rdma_umap_ops; 672 673 static int ib_uverbs_mmap(struct file *filp, struct vm_area_struct *vma) 674 { 675 struct ib_uverbs_file *file = filp->private_data; 676 struct ib_ucontext *ucontext; 677 int ret = 0; 678 int srcu_key; 679 680 srcu_key = srcu_read_lock(&file->device->disassociate_srcu); 681 ucontext = ib_uverbs_get_ucontext_file(file); 682 if (IS_ERR(ucontext)) { 683 ret = PTR_ERR(ucontext); 684 goto out; 685 } 686 vma->vm_ops = &rdma_umap_ops; 687 ret = ucontext->device->ops.mmap(ucontext, vma); 688 out: 689 srcu_read_unlock(&file->device->disassociate_srcu, srcu_key); 690 return ret; 691 } 692 693 /* 694 * The VMA has been dup'd, initialize the vm_private_data with a new tracking 695 * struct 696 */ 697 static void rdma_umap_open(struct vm_area_struct *vma) 698 { 699 struct ib_uverbs_file *ufile = vma->vm_file->private_data; 700 struct rdma_umap_priv *opriv = vma->vm_private_data; 701 struct rdma_umap_priv *priv; 702 703 if (!opriv) 704 return; 705 706 /* We are racing with disassociation */ 707 if (!down_read_trylock(&ufile->hw_destroy_rwsem)) 708 goto out_zap; 709 /* 710 * Disassociation already completed, the VMA should already be zapped. 711 */ 712 if (!ufile->ucontext) 713 goto out_unlock; 714 715 priv = kzalloc(sizeof(*priv), GFP_KERNEL); 716 if (!priv) 717 goto out_unlock; 718 rdma_umap_priv_init(priv, vma, opriv->entry); 719 720 up_read(&ufile->hw_destroy_rwsem); 721 return; 722 723 out_unlock: 724 up_read(&ufile->hw_destroy_rwsem); 725 out_zap: 726 /* 727 * We can't allow the VMA to be created with the actual IO pages, that 728 * would break our API contract, and it can't be stopped at this 729 * point, so zap it. 730 */ 731 vma->vm_private_data = NULL; 732 zap_vma_ptes(vma, vma->vm_start, vma->vm_end - vma->vm_start); 733 } 734 735 static void rdma_umap_close(struct vm_area_struct *vma) 736 { 737 struct ib_uverbs_file *ufile = vma->vm_file->private_data; 738 struct rdma_umap_priv *priv = vma->vm_private_data; 739 740 if (!priv) 741 return; 742 743 /* 744 * The vma holds a reference on the struct file that created it, which 745 * in turn means that the ib_uverbs_file is guaranteed to exist at 746 * this point. 747 */ 748 mutex_lock(&ufile->umap_lock); 749 if (priv->entry) 750 rdma_user_mmap_entry_put(priv->entry); 751 752 list_del(&priv->list); 753 mutex_unlock(&ufile->umap_lock); 754 kfree(priv); 755 } 756 757 /* 758 * Once the zap_vma_ptes has been called touches to the VMA will come here and 759 * we return a dummy writable zero page for all the pfns. 760 */ 761 static vm_fault_t rdma_umap_fault(struct vm_fault *vmf) 762 { 763 struct ib_uverbs_file *ufile = vmf->vma->vm_file->private_data; 764 struct rdma_umap_priv *priv = vmf->vma->vm_private_data; 765 vm_fault_t ret = 0; 766 767 if (!priv) 768 return VM_FAULT_SIGBUS; 769 770 /* Read only pages can just use the system zero page. */ 771 if (!(vmf->vma->vm_flags & (VM_WRITE | VM_MAYWRITE))) { 772 vmf->page = ZERO_PAGE(vmf->address); 773 get_page(vmf->page); 774 return 0; 775 } 776 777 mutex_lock(&ufile->umap_lock); 778 if (!ufile->disassociate_page) 779 ufile->disassociate_page = 780 alloc_pages(vmf->gfp_mask | __GFP_ZERO, 0); 781 782 if (ufile->disassociate_page) { 783 /* 784 * This VMA is forced to always be shared so this doesn't have 785 * to worry about COW. 786 */ 787 vmf->page = ufile->disassociate_page; 788 get_page(vmf->page); 789 } else { 790 ret = VM_FAULT_SIGBUS; 791 } 792 mutex_unlock(&ufile->umap_lock); 793 794 return ret; 795 } 796 797 static const struct vm_operations_struct rdma_umap_ops = { 798 .open = rdma_umap_open, 799 .close = rdma_umap_close, 800 .fault = rdma_umap_fault, 801 }; 802 803 void uverbs_user_mmap_disassociate(struct ib_uverbs_file *ufile) 804 { 805 struct rdma_umap_priv *priv, *next_priv; 806 807 lockdep_assert_held(&ufile->hw_destroy_rwsem); 808 809 while (1) { 810 struct mm_struct *mm = NULL; 811 812 /* Get an arbitrary mm pointer that hasn't been cleaned yet */ 813 mutex_lock(&ufile->umap_lock); 814 while (!list_empty(&ufile->umaps)) { 815 int ret; 816 817 priv = list_first_entry(&ufile->umaps, 818 struct rdma_umap_priv, list); 819 mm = priv->vma->vm_mm; 820 ret = mmget_not_zero(mm); 821 if (!ret) { 822 list_del_init(&priv->list); 823 mm = NULL; 824 continue; 825 } 826 break; 827 } 828 mutex_unlock(&ufile->umap_lock); 829 if (!mm) 830 return; 831 832 /* 833 * The umap_lock is nested under mmap_sem since it used within 834 * the vma_ops callbacks, so we have to clean the list one mm 835 * at a time to get the lock ordering right. Typically there 836 * will only be one mm, so no big deal. 837 */ 838 down_read(&mm->mmap_sem); 839 if (!mmget_still_valid(mm)) 840 goto skip_mm; 841 mutex_lock(&ufile->umap_lock); 842 list_for_each_entry_safe (priv, next_priv, &ufile->umaps, 843 list) { 844 struct vm_area_struct *vma = priv->vma; 845 846 if (vma->vm_mm != mm) 847 continue; 848 list_del_init(&priv->list); 849 850 zap_vma_ptes(vma, vma->vm_start, 851 vma->vm_end - vma->vm_start); 852 853 if (priv->entry) { 854 rdma_user_mmap_entry_put(priv->entry); 855 priv->entry = NULL; 856 } 857 } 858 mutex_unlock(&ufile->umap_lock); 859 skip_mm: 860 up_read(&mm->mmap_sem); 861 mmput(mm); 862 } 863 } 864 865 /* 866 * ib_uverbs_open() does not need the BKL: 867 * 868 * - the ib_uverbs_device structures are properly reference counted and 869 * everything else is purely local to the file being created, so 870 * races against other open calls are not a problem; 871 * - there is no ioctl method to race against; 872 * - the open method will either immediately run -ENXIO, or all 873 * required initialization will be done. 874 */ 875 static int ib_uverbs_open(struct inode *inode, struct file *filp) 876 { 877 struct ib_uverbs_device *dev; 878 struct ib_uverbs_file *file; 879 struct ib_device *ib_dev; 880 int ret; 881 int module_dependent; 882 int srcu_key; 883 884 dev = container_of(inode->i_cdev, struct ib_uverbs_device, cdev); 885 if (!atomic_inc_not_zero(&dev->refcount)) 886 return -ENXIO; 887 888 get_device(&dev->dev); 889 srcu_key = srcu_read_lock(&dev->disassociate_srcu); 890 mutex_lock(&dev->lists_mutex); 891 ib_dev = srcu_dereference(dev->ib_dev, 892 &dev->disassociate_srcu); 893 if (!ib_dev) { 894 ret = -EIO; 895 goto err; 896 } 897 898 if (!rdma_dev_access_netns(ib_dev, current->nsproxy->net_ns)) { 899 ret = -EPERM; 900 goto err; 901 } 902 903 /* In case IB device supports disassociate ucontext, there is no hard 904 * dependency between uverbs device and its low level device. 905 */ 906 module_dependent = !(ib_dev->ops.disassociate_ucontext); 907 908 if (module_dependent) { 909 if (!try_module_get(ib_dev->ops.owner)) { 910 ret = -ENODEV; 911 goto err; 912 } 913 } 914 915 file = kzalloc(sizeof(*file), GFP_KERNEL); 916 if (!file) { 917 ret = -ENOMEM; 918 if (module_dependent) 919 goto err_module; 920 921 goto err; 922 } 923 924 file->device = dev; 925 kref_init(&file->ref); 926 mutex_init(&file->ucontext_lock); 927 928 spin_lock_init(&file->uobjects_lock); 929 INIT_LIST_HEAD(&file->uobjects); 930 init_rwsem(&file->hw_destroy_rwsem); 931 mutex_init(&file->umap_lock); 932 INIT_LIST_HEAD(&file->umaps); 933 934 filp->private_data = file; 935 list_add_tail(&file->list, &dev->uverbs_file_list); 936 mutex_unlock(&dev->lists_mutex); 937 srcu_read_unlock(&dev->disassociate_srcu, srcu_key); 938 939 setup_ufile_idr_uobject(file); 940 941 return stream_open(inode, filp); 942 943 err_module: 944 module_put(ib_dev->ops.owner); 945 946 err: 947 mutex_unlock(&dev->lists_mutex); 948 srcu_read_unlock(&dev->disassociate_srcu, srcu_key); 949 if (atomic_dec_and_test(&dev->refcount)) 950 ib_uverbs_comp_dev(dev); 951 952 put_device(&dev->dev); 953 return ret; 954 } 955 956 static int ib_uverbs_close(struct inode *inode, struct file *filp) 957 { 958 struct ib_uverbs_file *file = filp->private_data; 959 960 uverbs_destroy_ufile_hw(file, RDMA_REMOVE_CLOSE); 961 962 mutex_lock(&file->device->lists_mutex); 963 list_del_init(&file->list); 964 mutex_unlock(&file->device->lists_mutex); 965 966 kref_put(&file->ref, ib_uverbs_release_file); 967 968 return 0; 969 } 970 971 static const struct file_operations uverbs_fops = { 972 .owner = THIS_MODULE, 973 .write = ib_uverbs_write, 974 .open = ib_uverbs_open, 975 .release = ib_uverbs_close, 976 .llseek = no_llseek, 977 .unlocked_ioctl = ib_uverbs_ioctl, 978 .compat_ioctl = compat_ptr_ioctl, 979 }; 980 981 static const struct file_operations uverbs_mmap_fops = { 982 .owner = THIS_MODULE, 983 .write = ib_uverbs_write, 984 .mmap = ib_uverbs_mmap, 985 .open = ib_uverbs_open, 986 .release = ib_uverbs_close, 987 .llseek = no_llseek, 988 .unlocked_ioctl = ib_uverbs_ioctl, 989 .compat_ioctl = compat_ptr_ioctl, 990 }; 991 992 static int ib_uverbs_get_nl_info(struct ib_device *ibdev, void *client_data, 993 struct ib_client_nl_info *res) 994 { 995 struct ib_uverbs_device *uverbs_dev = client_data; 996 int ret; 997 998 if (res->port != -1) 999 return -EINVAL; 1000 1001 res->abi = ibdev->ops.uverbs_abi_ver; 1002 res->cdev = &uverbs_dev->dev; 1003 1004 /* 1005 * To support DRIVER_ID binding in userspace some of the driver need 1006 * upgrading to expose their PCI dependent revision information 1007 * through get_context instead of relying on modalias matching. When 1008 * the drivers are fixed they can drop this flag. 1009 */ 1010 if (!ibdev->ops.uverbs_no_driver_id_binding) { 1011 ret = nla_put_u32(res->nl_msg, RDMA_NLDEV_ATTR_UVERBS_DRIVER_ID, 1012 ibdev->ops.driver_id); 1013 if (ret) 1014 return ret; 1015 } 1016 return 0; 1017 } 1018 1019 static struct ib_client uverbs_client = { 1020 .name = "uverbs", 1021 .no_kverbs_req = true, 1022 .add = ib_uverbs_add_one, 1023 .remove = ib_uverbs_remove_one, 1024 .get_nl_info = ib_uverbs_get_nl_info, 1025 }; 1026 MODULE_ALIAS_RDMA_CLIENT("uverbs"); 1027 1028 static ssize_t ibdev_show(struct device *device, struct device_attribute *attr, 1029 char *buf) 1030 { 1031 struct ib_uverbs_device *dev = 1032 container_of(device, struct ib_uverbs_device, dev); 1033 int ret = -ENODEV; 1034 int srcu_key; 1035 struct ib_device *ib_dev; 1036 1037 srcu_key = srcu_read_lock(&dev->disassociate_srcu); 1038 ib_dev = srcu_dereference(dev->ib_dev, &dev->disassociate_srcu); 1039 if (ib_dev) 1040 ret = sprintf(buf, "%s\n", dev_name(&ib_dev->dev)); 1041 srcu_read_unlock(&dev->disassociate_srcu, srcu_key); 1042 1043 return ret; 1044 } 1045 static DEVICE_ATTR_RO(ibdev); 1046 1047 static ssize_t abi_version_show(struct device *device, 1048 struct device_attribute *attr, char *buf) 1049 { 1050 struct ib_uverbs_device *dev = 1051 container_of(device, struct ib_uverbs_device, dev); 1052 int ret = -ENODEV; 1053 int srcu_key; 1054 struct ib_device *ib_dev; 1055 1056 srcu_key = srcu_read_lock(&dev->disassociate_srcu); 1057 ib_dev = srcu_dereference(dev->ib_dev, &dev->disassociate_srcu); 1058 if (ib_dev) 1059 ret = sprintf(buf, "%u\n", ib_dev->ops.uverbs_abi_ver); 1060 srcu_read_unlock(&dev->disassociate_srcu, srcu_key); 1061 1062 return ret; 1063 } 1064 static DEVICE_ATTR_RO(abi_version); 1065 1066 static struct attribute *ib_dev_attrs[] = { 1067 &dev_attr_abi_version.attr, 1068 &dev_attr_ibdev.attr, 1069 NULL, 1070 }; 1071 1072 static const struct attribute_group dev_attr_group = { 1073 .attrs = ib_dev_attrs, 1074 }; 1075 1076 static CLASS_ATTR_STRING(abi_version, S_IRUGO, 1077 __stringify(IB_USER_VERBS_ABI_VERSION)); 1078 1079 static int ib_uverbs_create_uapi(struct ib_device *device, 1080 struct ib_uverbs_device *uverbs_dev) 1081 { 1082 struct uverbs_api *uapi; 1083 1084 uapi = uverbs_alloc_api(device); 1085 if (IS_ERR(uapi)) 1086 return PTR_ERR(uapi); 1087 1088 uverbs_dev->uapi = uapi; 1089 return 0; 1090 } 1091 1092 static void ib_uverbs_add_one(struct ib_device *device) 1093 { 1094 int devnum; 1095 dev_t base; 1096 struct ib_uverbs_device *uverbs_dev; 1097 int ret; 1098 1099 if (!device->ops.alloc_ucontext) 1100 return; 1101 1102 uverbs_dev = kzalloc(sizeof(*uverbs_dev), GFP_KERNEL); 1103 if (!uverbs_dev) 1104 return; 1105 1106 ret = init_srcu_struct(&uverbs_dev->disassociate_srcu); 1107 if (ret) { 1108 kfree(uverbs_dev); 1109 return; 1110 } 1111 1112 device_initialize(&uverbs_dev->dev); 1113 uverbs_dev->dev.class = uverbs_class; 1114 uverbs_dev->dev.parent = device->dev.parent; 1115 uverbs_dev->dev.release = ib_uverbs_release_dev; 1116 uverbs_dev->groups[0] = &dev_attr_group; 1117 uverbs_dev->dev.groups = uverbs_dev->groups; 1118 atomic_set(&uverbs_dev->refcount, 1); 1119 init_completion(&uverbs_dev->comp); 1120 uverbs_dev->xrcd_tree = RB_ROOT; 1121 mutex_init(&uverbs_dev->xrcd_tree_mutex); 1122 mutex_init(&uverbs_dev->lists_mutex); 1123 INIT_LIST_HEAD(&uverbs_dev->uverbs_file_list); 1124 rcu_assign_pointer(uverbs_dev->ib_dev, device); 1125 uverbs_dev->num_comp_vectors = device->num_comp_vectors; 1126 1127 devnum = ida_alloc_max(&uverbs_ida, IB_UVERBS_MAX_DEVICES - 1, 1128 GFP_KERNEL); 1129 if (devnum < 0) 1130 goto err; 1131 uverbs_dev->devnum = devnum; 1132 if (devnum >= IB_UVERBS_NUM_FIXED_MINOR) 1133 base = dynamic_uverbs_dev + devnum - IB_UVERBS_NUM_FIXED_MINOR; 1134 else 1135 base = IB_UVERBS_BASE_DEV + devnum; 1136 1137 if (ib_uverbs_create_uapi(device, uverbs_dev)) 1138 goto err_uapi; 1139 1140 uverbs_dev->dev.devt = base; 1141 dev_set_name(&uverbs_dev->dev, "uverbs%d", uverbs_dev->devnum); 1142 1143 cdev_init(&uverbs_dev->cdev, 1144 device->ops.mmap ? &uverbs_mmap_fops : &uverbs_fops); 1145 uverbs_dev->cdev.owner = THIS_MODULE; 1146 1147 ret = cdev_device_add(&uverbs_dev->cdev, &uverbs_dev->dev); 1148 if (ret) 1149 goto err_uapi; 1150 1151 ib_set_client_data(device, &uverbs_client, uverbs_dev); 1152 return; 1153 1154 err_uapi: 1155 ida_free(&uverbs_ida, devnum); 1156 err: 1157 if (atomic_dec_and_test(&uverbs_dev->refcount)) 1158 ib_uverbs_comp_dev(uverbs_dev); 1159 wait_for_completion(&uverbs_dev->comp); 1160 put_device(&uverbs_dev->dev); 1161 return; 1162 } 1163 1164 static void ib_uverbs_free_hw_resources(struct ib_uverbs_device *uverbs_dev, 1165 struct ib_device *ib_dev) 1166 { 1167 struct ib_uverbs_file *file; 1168 1169 /* Pending running commands to terminate */ 1170 uverbs_disassociate_api_pre(uverbs_dev); 1171 1172 mutex_lock(&uverbs_dev->lists_mutex); 1173 while (!list_empty(&uverbs_dev->uverbs_file_list)) { 1174 file = list_first_entry(&uverbs_dev->uverbs_file_list, 1175 struct ib_uverbs_file, list); 1176 list_del_init(&file->list); 1177 kref_get(&file->ref); 1178 1179 /* We must release the mutex before going ahead and calling 1180 * uverbs_cleanup_ufile, as it might end up indirectly calling 1181 * uverbs_close, for example due to freeing the resources (e.g 1182 * mmput). 1183 */ 1184 mutex_unlock(&uverbs_dev->lists_mutex); 1185 1186 ib_uverbs_async_handler(READ_ONCE(file->async_file), 0, 1187 IB_EVENT_DEVICE_FATAL, NULL, NULL); 1188 1189 uverbs_destroy_ufile_hw(file, RDMA_REMOVE_DRIVER_REMOVE); 1190 kref_put(&file->ref, ib_uverbs_release_file); 1191 1192 mutex_lock(&uverbs_dev->lists_mutex); 1193 } 1194 mutex_unlock(&uverbs_dev->lists_mutex); 1195 1196 uverbs_disassociate_api(uverbs_dev->uapi); 1197 } 1198 1199 static void ib_uverbs_remove_one(struct ib_device *device, void *client_data) 1200 { 1201 struct ib_uverbs_device *uverbs_dev = client_data; 1202 int wait_clients = 1; 1203 1204 if (!uverbs_dev) 1205 return; 1206 1207 cdev_device_del(&uverbs_dev->cdev, &uverbs_dev->dev); 1208 ida_free(&uverbs_ida, uverbs_dev->devnum); 1209 1210 if (device->ops.disassociate_ucontext) { 1211 /* We disassociate HW resources and immediately return. 1212 * Userspace will see a EIO errno for all future access. 1213 * Upon returning, ib_device may be freed internally and is not 1214 * valid any more. 1215 * uverbs_device is still available until all clients close 1216 * their files, then the uverbs device ref count will be zero 1217 * and its resources will be freed. 1218 * Note: At this point no more files can be opened since the 1219 * cdev was deleted, however active clients can still issue 1220 * commands and close their open files. 1221 */ 1222 ib_uverbs_free_hw_resources(uverbs_dev, device); 1223 wait_clients = 0; 1224 } 1225 1226 if (atomic_dec_and_test(&uverbs_dev->refcount)) 1227 ib_uverbs_comp_dev(uverbs_dev); 1228 if (wait_clients) 1229 wait_for_completion(&uverbs_dev->comp); 1230 1231 put_device(&uverbs_dev->dev); 1232 } 1233 1234 static char *uverbs_devnode(struct device *dev, umode_t *mode) 1235 { 1236 if (mode) 1237 *mode = 0666; 1238 return kasprintf(GFP_KERNEL, "infiniband/%s", dev_name(dev)); 1239 } 1240 1241 static int __init ib_uverbs_init(void) 1242 { 1243 int ret; 1244 1245 ret = register_chrdev_region(IB_UVERBS_BASE_DEV, 1246 IB_UVERBS_NUM_FIXED_MINOR, 1247 "infiniband_verbs"); 1248 if (ret) { 1249 pr_err("user_verbs: couldn't register device number\n"); 1250 goto out; 1251 } 1252 1253 ret = alloc_chrdev_region(&dynamic_uverbs_dev, 0, 1254 IB_UVERBS_NUM_DYNAMIC_MINOR, 1255 "infiniband_verbs"); 1256 if (ret) { 1257 pr_err("couldn't register dynamic device number\n"); 1258 goto out_alloc; 1259 } 1260 1261 uverbs_class = class_create(THIS_MODULE, "infiniband_verbs"); 1262 if (IS_ERR(uverbs_class)) { 1263 ret = PTR_ERR(uverbs_class); 1264 pr_err("user_verbs: couldn't create class infiniband_verbs\n"); 1265 goto out_chrdev; 1266 } 1267 1268 uverbs_class->devnode = uverbs_devnode; 1269 1270 ret = class_create_file(uverbs_class, &class_attr_abi_version.attr); 1271 if (ret) { 1272 pr_err("user_verbs: couldn't create abi_version attribute\n"); 1273 goto out_class; 1274 } 1275 1276 ret = ib_register_client(&uverbs_client); 1277 if (ret) { 1278 pr_err("user_verbs: couldn't register client\n"); 1279 goto out_class; 1280 } 1281 1282 return 0; 1283 1284 out_class: 1285 class_destroy(uverbs_class); 1286 1287 out_chrdev: 1288 unregister_chrdev_region(dynamic_uverbs_dev, 1289 IB_UVERBS_NUM_DYNAMIC_MINOR); 1290 1291 out_alloc: 1292 unregister_chrdev_region(IB_UVERBS_BASE_DEV, 1293 IB_UVERBS_NUM_FIXED_MINOR); 1294 1295 out: 1296 return ret; 1297 } 1298 1299 static void __exit ib_uverbs_cleanup(void) 1300 { 1301 ib_unregister_client(&uverbs_client); 1302 class_destroy(uverbs_class); 1303 unregister_chrdev_region(IB_UVERBS_BASE_DEV, 1304 IB_UVERBS_NUM_FIXED_MINOR); 1305 unregister_chrdev_region(dynamic_uverbs_dev, 1306 IB_UVERBS_NUM_DYNAMIC_MINOR); 1307 mmu_notifier_synchronize(); 1308 } 1309 1310 module_init(ib_uverbs_init); 1311 module_exit(ib_uverbs_cleanup); 1312