1 /*- 2 * SPDX-License-Identifier: BSD-2-Clause OR GPL-2.0 3 * 4 * Copyright (c) 2005 Topspin Communications. All rights reserved. 5 * Copyright (c) 2005, 2006 Cisco Systems. All rights reserved. 6 * Copyright (c) 2005 Mellanox Technologies. All rights reserved. 7 * Copyright (c) 2005 Voltaire, Inc. All rights reserved. 8 * Copyright (c) 2005 PathScale, Inc. All rights reserved. 9 * 10 * This software is available to you under a choice of one of two 11 * licenses. You may choose to be licensed under the terms of the GNU 12 * General Public License (GPL) Version 2, available from the file 13 * COPYING in the main directory of this source tree, or the 14 * OpenIB.org BSD license below: 15 * 16 * Redistribution and use in source and binary forms, with or 17 * without modification, are permitted provided that the following 18 * conditions are met: 19 * 20 * - Redistributions of source code must retain the above 21 * copyright notice, this list of conditions and the following 22 * disclaimer. 23 * 24 * - Redistributions in binary form must reproduce the above 25 * copyright notice, this list of conditions and the following 26 * disclaimer in the documentation and/or other materials 27 * provided with the distribution. 28 * 29 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, 30 * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF 31 * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND 32 * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS 33 * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN 34 * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN 35 * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE 36 * SOFTWARE. 37 */ 38 39 #include <sys/cdefs.h> 40 __FBSDID("$FreeBSD$"); 41 42 #include <linux/module.h> 43 #include <linux/device.h> 44 #include <linux/err.h> 45 #include <linux/fs.h> 46 #include <linux/poll.h> 47 #include <linux/sched.h> 48 #include <linux/file.h> 49 #include <linux/cdev.h> 50 #include <linux/slab.h> 51 #include <linux/pci.h> 52 #include <linux/lockdep.h> 53 54 #include <asm/uaccess.h> 55 56 #include <rdma/ib.h> 57 #include <rdma/uverbs_std_types.h> 58 59 #include "uverbs.h" 60 #include "core_priv.h" 61 #include "rdma_core.h" 62 63 MODULE_AUTHOR("Roland Dreier"); 64 MODULE_DESCRIPTION("InfiniBand userspace verbs access"); 65 MODULE_LICENSE("Dual BSD/GPL"); 66 67 enum { 68 IB_UVERBS_MAJOR = 231, 69 IB_UVERBS_BASE_MINOR = 192, 70 IB_UVERBS_MAX_DEVICES = RDMA_MAX_PORTS, 71 IB_UVERBS_NUM_FIXED_MINOR = 32, 72 IB_UVERBS_NUM_DYNAMIC_MINOR = IB_UVERBS_MAX_DEVICES - IB_UVERBS_NUM_FIXED_MINOR, 73 }; 74 75 #define IB_UVERBS_BASE_DEV MKDEV(IB_UVERBS_MAJOR, IB_UVERBS_BASE_MINOR) 76 77 static dev_t dynamic_uverbs_dev; 78 static struct class *uverbs_class; 79 80 static DEFINE_IDA(uverbs_ida); 81 static void ib_uverbs_add_one(struct ib_device *device); 82 static void ib_uverbs_remove_one(struct ib_device *device, void *client_data); 83 84 /* 85 * Must be called with the ufile->device->disassociate_srcu held, and the lock 86 * must be held until use of the ucontext is finished. 87 */ 88 struct ib_ucontext *ib_uverbs_get_ucontext_file(struct ib_uverbs_file *ufile) 89 { 90 /* 91 * We do not hold the hw_destroy_rwsem lock for this flow, instead 92 * srcu is used. It does not matter if someone races this with 93 * get_context, we get NULL or valid ucontext. 94 */ 95 struct ib_ucontext *ucontext = READ_ONCE(ufile->ucontext); 96 97 if (!srcu_dereference(ufile->device->ib_dev, 98 &ufile->device->disassociate_srcu)) 99 return ERR_PTR(-EIO); 100 101 if (!ucontext) 102 return ERR_PTR(-EINVAL); 103 104 return ucontext; 105 } 106 EXPORT_SYMBOL(ib_uverbs_get_ucontext_file); 107 108 int uverbs_dealloc_mw(struct ib_mw *mw) 109 { 110 struct ib_pd *pd = mw->pd; 111 int ret; 112 113 ret = mw->device->dealloc_mw(mw); 114 if (!ret) 115 atomic_dec(&pd->usecnt); 116 return ret; 117 } 118 119 static void ib_uverbs_release_dev(struct device *device) 120 { 121 struct ib_uverbs_device *dev = 122 container_of(device, struct ib_uverbs_device, dev); 123 124 uverbs_destroy_api(dev->uapi); 125 cleanup_srcu_struct(&dev->disassociate_srcu); 126 mutex_destroy(&dev->lists_mutex); 127 mutex_destroy(&dev->xrcd_tree_mutex); 128 kfree(dev); 129 } 130 131 void ib_uverbs_release_ucq(struct ib_uverbs_completion_event_file *ev_file, 132 struct ib_ucq_object *uobj) 133 { 134 struct ib_uverbs_event *evt, *tmp; 135 136 if (ev_file) { 137 spin_lock_irq(&ev_file->ev_queue.lock); 138 list_for_each_entry_safe(evt, tmp, &uobj->comp_list, obj_list) { 139 list_del(&evt->list); 140 kfree(evt); 141 } 142 spin_unlock_irq(&ev_file->ev_queue.lock); 143 144 uverbs_uobject_put(&ev_file->uobj); 145 } 146 147 ib_uverbs_release_uevent(&uobj->uevent); 148 } 149 150 void ib_uverbs_release_uevent(struct ib_uevent_object *uobj) 151 { 152 struct ib_uverbs_async_event_file *async_file = 153 READ_ONCE(uobj->uobject.ufile->async_file); 154 struct ib_uverbs_event *evt, *tmp; 155 156 if (!async_file) 157 return; 158 159 spin_lock_irq(&async_file->ev_queue.lock); 160 list_for_each_entry_safe(evt, tmp, &uobj->event_list, obj_list) { 161 list_del(&evt->list); 162 kfree(evt); 163 } 164 spin_unlock_irq(&async_file->ev_queue.lock); 165 } 166 167 void ib_uverbs_detach_umcast(struct ib_qp *qp, 168 struct ib_uqp_object *uobj) 169 { 170 struct ib_uverbs_mcast_entry *mcast, *tmp; 171 172 list_for_each_entry_safe(mcast, tmp, &uobj->mcast_list, list) { 173 ib_detach_mcast(qp, &mcast->gid, mcast->lid); 174 list_del(&mcast->list); 175 kfree(mcast); 176 } 177 } 178 179 static void ib_uverbs_comp_dev(struct ib_uverbs_device *dev) 180 { 181 complete(&dev->comp); 182 } 183 184 void ib_uverbs_release_file(struct kref *ref) 185 { 186 struct ib_uverbs_file *file = 187 container_of(ref, struct ib_uverbs_file, ref); 188 struct ib_device *ib_dev; 189 int srcu_key; 190 191 release_ufile_idr_uobject(file); 192 193 srcu_key = srcu_read_lock(&file->device->disassociate_srcu); 194 ib_dev = srcu_dereference(file->device->ib_dev, 195 &file->device->disassociate_srcu); 196 if (ib_dev && !ib_dev->disassociate_ucontext) 197 module_put(ib_dev->owner); 198 srcu_read_unlock(&file->device->disassociate_srcu, srcu_key); 199 200 if (atomic_dec_and_test(&file->device->refcount)) 201 ib_uverbs_comp_dev(file->device); 202 203 if (file->async_file) 204 uverbs_uobject_put(&file->async_file->uobj); 205 put_device(&file->device->dev); 206 207 mutex_destroy(&file->umap_lock); 208 mutex_destroy(&file->ucontext_lock); 209 kfree(file); 210 } 211 212 static ssize_t ib_uverbs_event_read(struct ib_uverbs_event_queue *ev_queue, 213 struct file *filp, char __user *buf, 214 size_t count, loff_t *pos, 215 size_t eventsz) 216 { 217 struct ib_uverbs_event *event; 218 int ret = 0; 219 220 spin_lock_irq(&ev_queue->lock); 221 222 while (list_empty(&ev_queue->event_list)) { 223 spin_unlock_irq(&ev_queue->lock); 224 225 if (filp->f_flags & O_NONBLOCK) 226 return -EAGAIN; 227 228 if (wait_event_interruptible(ev_queue->poll_wait, 229 (!list_empty(&ev_queue->event_list) || 230 ev_queue->is_closed))) 231 return -ERESTARTSYS; 232 233 spin_lock_irq(&ev_queue->lock); 234 235 /* If device was disassociated and no event exists set an error */ 236 if (list_empty(&ev_queue->event_list) && ev_queue->is_closed) { 237 spin_unlock_irq(&ev_queue->lock); 238 return -EIO; 239 } 240 } 241 242 event = list_entry(ev_queue->event_list.next, struct ib_uverbs_event, list); 243 244 if (eventsz > count) { 245 ret = -EINVAL; 246 event = NULL; 247 } else { 248 list_del(ev_queue->event_list.next); 249 if (event->counter) { 250 ++(*event->counter); 251 list_del(&event->obj_list); 252 } 253 } 254 255 spin_unlock_irq(&ev_queue->lock); 256 257 if (event) { 258 if (copy_to_user(buf, event, eventsz)) 259 ret = -EFAULT; 260 else 261 ret = eventsz; 262 } 263 264 kfree(event); 265 266 return ret; 267 } 268 269 static ssize_t ib_uverbs_async_event_read(struct file *filp, char __user *buf, 270 size_t count, loff_t *pos) 271 { 272 struct ib_uverbs_async_event_file *file = filp->private_data; 273 274 return ib_uverbs_event_read(&file->ev_queue, filp, buf, count, pos, 275 sizeof(struct ib_uverbs_async_event_desc)); 276 } 277 278 static ssize_t ib_uverbs_comp_event_read(struct file *filp, char __user *buf, 279 size_t count, loff_t *pos) 280 { 281 struct ib_uverbs_completion_event_file *comp_ev_file = 282 filp->private_data; 283 284 return ib_uverbs_event_read(&comp_ev_file->ev_queue, filp, buf, count, 285 pos, 286 sizeof(struct ib_uverbs_comp_event_desc)); 287 } 288 289 static __poll_t ib_uverbs_event_poll(struct ib_uverbs_event_queue *ev_queue, 290 struct file *filp, 291 struct poll_table_struct *wait) 292 { 293 __poll_t pollflags = 0; 294 295 poll_wait(filp, &ev_queue->poll_wait, wait); 296 297 spin_lock_irq(&ev_queue->lock); 298 if (!list_empty(&ev_queue->event_list)) 299 pollflags = POLLIN | POLLRDNORM; 300 spin_unlock_irq(&ev_queue->lock); 301 302 return pollflags; 303 } 304 305 static __poll_t ib_uverbs_async_event_poll(struct file *filp, 306 struct poll_table_struct *wait) 307 { 308 struct ib_uverbs_async_event_file *file = filp->private_data; 309 310 return ib_uverbs_event_poll(&file->ev_queue, filp, wait); 311 } 312 313 static __poll_t ib_uverbs_comp_event_poll(struct file *filp, 314 struct poll_table_struct *wait) 315 { 316 struct ib_uverbs_completion_event_file *comp_ev_file = 317 filp->private_data; 318 319 return ib_uverbs_event_poll(&comp_ev_file->ev_queue, filp, wait); 320 } 321 322 static int ib_uverbs_async_event_fasync(int fd, struct file *filp, int on) 323 { 324 struct ib_uverbs_async_event_file *file = filp->private_data; 325 326 return fasync_helper(fd, filp, on, &file->ev_queue.async_queue); 327 } 328 329 static int ib_uverbs_comp_event_fasync(int fd, struct file *filp, int on) 330 { 331 struct ib_uverbs_completion_event_file *comp_ev_file = 332 filp->private_data; 333 334 return fasync_helper(fd, filp, on, &comp_ev_file->ev_queue.async_queue); 335 } 336 337 const struct file_operations uverbs_event_fops = { 338 .owner = THIS_MODULE, 339 .read = ib_uverbs_comp_event_read, 340 .poll = ib_uverbs_comp_event_poll, 341 .release = uverbs_uobject_fd_release, 342 .fasync = ib_uverbs_comp_event_fasync, 343 .llseek = no_llseek, 344 }; 345 346 const struct file_operations uverbs_async_event_fops = { 347 .owner = THIS_MODULE, 348 .read = ib_uverbs_async_event_read, 349 .poll = ib_uverbs_async_event_poll, 350 .release = uverbs_uobject_fd_release, 351 .fasync = ib_uverbs_async_event_fasync, 352 .llseek = no_llseek, 353 }; 354 355 void ib_uverbs_comp_handler(struct ib_cq *cq, void *cq_context) 356 { 357 struct ib_uverbs_event_queue *ev_queue = cq_context; 358 struct ib_ucq_object *uobj; 359 struct ib_uverbs_event *entry; 360 unsigned long flags; 361 362 if (!ev_queue) 363 return; 364 365 spin_lock_irqsave(&ev_queue->lock, flags); 366 if (ev_queue->is_closed) { 367 spin_unlock_irqrestore(&ev_queue->lock, flags); 368 return; 369 } 370 371 entry = kmalloc(sizeof(*entry), GFP_ATOMIC); 372 if (!entry) { 373 spin_unlock_irqrestore(&ev_queue->lock, flags); 374 return; 375 } 376 377 uobj = cq->uobject; 378 379 entry->desc.comp.cq_handle = cq->uobject->uevent.uobject.user_handle; 380 entry->counter = &uobj->comp_events_reported; 381 382 list_add_tail(&entry->list, &ev_queue->event_list); 383 list_add_tail(&entry->obj_list, &uobj->comp_list); 384 spin_unlock_irqrestore(&ev_queue->lock, flags); 385 386 wake_up_interruptible(&ev_queue->poll_wait); 387 kill_fasync(&ev_queue->async_queue, SIGIO, POLL_IN); 388 } 389 390 static void 391 ib_uverbs_async_handler(struct ib_uverbs_async_event_file *async_file, 392 __u64 element, __u64 event, struct list_head *obj_list, 393 u32 *counter) 394 { 395 struct ib_uverbs_event *entry; 396 unsigned long flags; 397 398 if (!async_file) 399 return; 400 401 spin_lock_irqsave(&async_file->ev_queue.lock, flags); 402 if (async_file->ev_queue.is_closed) { 403 spin_unlock_irqrestore(&async_file->ev_queue.lock, flags); 404 return; 405 } 406 407 entry = kmalloc(sizeof(*entry), GFP_ATOMIC); 408 if (!entry) { 409 spin_unlock_irqrestore(&async_file->ev_queue.lock, flags); 410 return; 411 } 412 413 entry->desc.async.element = element; 414 entry->desc.async.event_type = event; 415 entry->desc.async.reserved = 0; 416 entry->counter = counter; 417 418 list_add_tail(&entry->list, &async_file->ev_queue.event_list); 419 if (obj_list) 420 list_add_tail(&entry->obj_list, obj_list); 421 spin_unlock_irqrestore(&async_file->ev_queue.lock, flags); 422 423 wake_up_interruptible(&async_file->ev_queue.poll_wait); 424 kill_fasync(&async_file->ev_queue.async_queue, SIGIO, POLL_IN); 425 } 426 427 static void uverbs_uobj_event(struct ib_uevent_object *eobj, 428 struct ib_event *event) 429 { 430 ib_uverbs_async_handler(READ_ONCE(eobj->uobject.ufile->async_file), 431 eobj->uobject.user_handle, event->event, 432 &eobj->event_list, &eobj->events_reported); 433 } 434 435 void ib_uverbs_cq_event_handler(struct ib_event *event, void *context_ptr) 436 { 437 uverbs_uobj_event(&event->element.cq->uobject->uevent, event); 438 } 439 440 void ib_uverbs_qp_event_handler(struct ib_event *event, void *context_ptr) 441 { 442 /* for XRC target qp's, check that qp is live */ 443 if (!event->element.qp->uobject) 444 return; 445 446 uverbs_uobj_event(&event->element.qp->uobject->uevent, event); 447 } 448 449 void ib_uverbs_wq_event_handler(struct ib_event *event, void *context_ptr) 450 { 451 uverbs_uobj_event(&event->element.wq->uobject->uevent, event); 452 } 453 454 void ib_uverbs_srq_event_handler(struct ib_event *event, void *context_ptr) 455 { 456 uverbs_uobj_event(&event->element.srq->uobject->uevent, event); 457 } 458 459 static void ib_uverbs_event_handler(struct ib_event_handler *handler, 460 struct ib_event *event) 461 { 462 ib_uverbs_async_handler( 463 container_of(handler, struct ib_uverbs_async_event_file, 464 event_handler), 465 event->element.port_num, event->event, NULL, NULL); 466 } 467 468 void ib_uverbs_init_event_queue(struct ib_uverbs_event_queue *ev_queue) 469 { 470 spin_lock_init(&ev_queue->lock); 471 INIT_LIST_HEAD(&ev_queue->event_list); 472 init_waitqueue_head(&ev_queue->poll_wait); 473 ev_queue->is_closed = 0; 474 ev_queue->async_queue = NULL; 475 } 476 477 void ib_uverbs_init_async_event_file( 478 struct ib_uverbs_async_event_file *async_file) 479 { 480 struct ib_uverbs_file *uverbs_file = async_file->uobj.ufile; 481 struct ib_device *ib_dev = async_file->uobj.context->device; 482 483 ib_uverbs_init_event_queue(&async_file->ev_queue); 484 485 /* The first async_event_file becomes the default one for the file. */ 486 mutex_lock(&uverbs_file->ucontext_lock); 487 if (!uverbs_file->async_file) { 488 /* Pairs with the put in ib_uverbs_release_file */ 489 uverbs_uobject_get(&async_file->uobj); 490 atomic_store_rel_ptr((uintptr_t *)&uverbs_file->async_file, (uintptr_t)async_file); 491 } 492 mutex_unlock(&uverbs_file->ucontext_lock); 493 494 INIT_IB_EVENT_HANDLER(&async_file->event_handler, ib_dev, 495 ib_uverbs_event_handler); 496 ib_register_event_handler(&async_file->event_handler); 497 } 498 499 static ssize_t verify_hdr(struct ib_uverbs_cmd_hdr *hdr, 500 struct ib_uverbs_ex_cmd_hdr *ex_hdr, size_t count, 501 const struct uverbs_api_write_method *method_elm) 502 { 503 if (method_elm->is_ex) { 504 count -= sizeof(*hdr) + sizeof(*ex_hdr); 505 506 if ((hdr->in_words + ex_hdr->provider_in_words) * 8 != count) 507 return -EINVAL; 508 509 if (hdr->in_words * 8 < method_elm->req_size) 510 return -ENOSPC; 511 512 if (ex_hdr->cmd_hdr_reserved) 513 return -EINVAL; 514 515 if (ex_hdr->response) { 516 if (!hdr->out_words && !ex_hdr->provider_out_words) 517 return -EINVAL; 518 519 if (hdr->out_words * 8 < method_elm->resp_size) 520 return -ENOSPC; 521 522 if (!access_ok(u64_to_user_ptr(ex_hdr->response), 523 (hdr->out_words + ex_hdr->provider_out_words) * 8)) 524 return -EFAULT; 525 } else { 526 if (hdr->out_words || ex_hdr->provider_out_words) 527 return -EINVAL; 528 } 529 530 return 0; 531 } 532 533 /* not extended command */ 534 if (hdr->in_words * 4 != count) 535 return -EINVAL; 536 537 if (count < method_elm->req_size + sizeof(hdr)) { 538 /* 539 * rdma-core v18 and v19 have a bug where they send DESTROY_CQ 540 * with a 16 byte write instead of 24. Old kernels didn't 541 * check the size so they allowed this. Now that the size is 542 * checked provide a compatibility work around to not break 543 * those userspaces. 544 */ 545 if (hdr->command == IB_USER_VERBS_CMD_DESTROY_CQ && 546 count == 16) { 547 hdr->in_words = 6; 548 return 0; 549 } 550 return -ENOSPC; 551 } 552 if (hdr->out_words * 4 < method_elm->resp_size) 553 return -ENOSPC; 554 555 return 0; 556 } 557 558 static ssize_t ib_uverbs_write(struct file *filp, const char __user *buf, 559 size_t count, loff_t *pos) 560 { 561 struct ib_uverbs_file *file = filp->private_data; 562 const struct uverbs_api_write_method *method_elm; 563 struct uverbs_api *uapi = file->device->uapi; 564 struct ib_uverbs_ex_cmd_hdr ex_hdr; 565 struct ib_uverbs_cmd_hdr hdr; 566 struct uverbs_attr_bundle bundle; 567 int srcu_key; 568 ssize_t ret; 569 570 if (!ib_safe_file_access(filp)) { 571 pr_warn_once("uverbs_write: process %d (%s) changed security contexts after opening file descriptor, this is not allowed.\n", 572 current->pid, current->comm); 573 return -EACCES; 574 } 575 576 if (count < sizeof(hdr)) 577 return -EINVAL; 578 579 if (copy_from_user(&hdr, buf, sizeof(hdr))) 580 return -EFAULT; 581 582 method_elm = uapi_get_method(uapi, hdr.command); 583 if (IS_ERR(method_elm)) 584 return PTR_ERR(method_elm); 585 586 if (method_elm->is_ex) { 587 if (count < (sizeof(hdr) + sizeof(ex_hdr))) 588 return -EINVAL; 589 if (copy_from_user(&ex_hdr, buf + sizeof(hdr), sizeof(ex_hdr))) 590 return -EFAULT; 591 } 592 593 ret = verify_hdr(&hdr, &ex_hdr, count, method_elm); 594 if (ret) 595 return ret; 596 597 srcu_key = srcu_read_lock(&file->device->disassociate_srcu); 598 599 buf += sizeof(hdr); 600 601 memset(bundle.attr_present, 0, sizeof(bundle.attr_present)); 602 bundle.ufile = file; 603 bundle.context = NULL; /* only valid if bundle has uobject */ 604 if (!method_elm->is_ex) { 605 size_t in_len = hdr.in_words * 4 - sizeof(hdr); 606 size_t out_len = hdr.out_words * 4; 607 u64 response = 0; 608 609 if (method_elm->has_udata) { 610 bundle.driver_udata.inlen = 611 in_len - method_elm->req_size; 612 in_len = method_elm->req_size; 613 if (bundle.driver_udata.inlen) 614 bundle.driver_udata.inbuf = buf + in_len; 615 else 616 bundle.driver_udata.inbuf = NULL; 617 } else { 618 memset(&bundle.driver_udata, 0, 619 sizeof(bundle.driver_udata)); 620 } 621 622 if (method_elm->has_resp) { 623 /* 624 * The macros check that if has_resp is set 625 * then the command request structure starts 626 * with a '__aligned u64 response' member. 627 */ 628 ret = get_user(response, (const u64 __user *)buf); 629 if (ret) 630 goto out_unlock; 631 632 if (method_elm->has_udata) { 633 bundle.driver_udata.outlen = 634 out_len - method_elm->resp_size; 635 out_len = method_elm->resp_size; 636 if (bundle.driver_udata.outlen) 637 bundle.driver_udata.outbuf = 638 u64_to_user_ptr(response + 639 out_len); 640 else 641 bundle.driver_udata.outbuf = NULL; 642 } 643 } else { 644 bundle.driver_udata.outlen = 0; 645 bundle.driver_udata.outbuf = NULL; 646 } 647 648 ib_uverbs_init_udata_buf_or_null( 649 &bundle.ucore, buf, u64_to_user_ptr(response), 650 in_len, out_len); 651 } else { 652 buf += sizeof(ex_hdr); 653 654 ib_uverbs_init_udata_buf_or_null(&bundle.ucore, buf, 655 u64_to_user_ptr(ex_hdr.response), 656 hdr.in_words * 8, hdr.out_words * 8); 657 658 ib_uverbs_init_udata_buf_or_null( 659 &bundle.driver_udata, buf + bundle.ucore.inlen, 660 u64_to_user_ptr(ex_hdr.response + bundle.ucore.outlen), 661 ex_hdr.provider_in_words * 8, 662 ex_hdr.provider_out_words * 8); 663 664 } 665 666 ret = method_elm->handler(&bundle); 667 out_unlock: 668 srcu_read_unlock(&file->device->disassociate_srcu, srcu_key); 669 return (ret) ? : count; 670 } 671 672 static const struct vm_operations_struct rdma_umap_ops; 673 674 static int ib_uverbs_mmap(struct file *filp, struct vm_area_struct *vma) 675 { 676 struct ib_uverbs_file *file = filp->private_data; 677 struct ib_ucontext *ucontext; 678 int ret = 0; 679 int srcu_key; 680 681 srcu_key = srcu_read_lock(&file->device->disassociate_srcu); 682 ucontext = ib_uverbs_get_ucontext_file(file); 683 if (IS_ERR(ucontext)) { 684 ret = PTR_ERR(ucontext); 685 goto out; 686 } 687 vma->vm_ops = &rdma_umap_ops; 688 ret = ucontext->device->mmap(ucontext, vma); 689 out: 690 srcu_read_unlock(&file->device->disassociate_srcu, srcu_key); 691 return ret; 692 } 693 694 /* 695 * The VMA has been dup'd, initialize the vm_private_data with a new tracking 696 * struct 697 */ 698 static void rdma_umap_open(struct vm_area_struct *vma) 699 { 700 struct ib_uverbs_file *ufile = vma->vm_file->private_data; 701 struct rdma_umap_priv *opriv = vma->vm_private_data; 702 struct rdma_umap_priv *priv; 703 704 if (!opriv) 705 return; 706 707 /* We are racing with disassociation */ 708 if (!down_read_trylock(&ufile->hw_destroy_rwsem)) 709 goto out_zap; 710 /* 711 * Disassociation already completed, the VMA should already be zapped. 712 */ 713 if (!ufile->ucontext) 714 goto out_unlock; 715 716 priv = kzalloc(sizeof(*priv), GFP_KERNEL); 717 if (!priv) 718 goto out_unlock; 719 rdma_umap_priv_init(priv, vma, opriv->entry); 720 721 up_read(&ufile->hw_destroy_rwsem); 722 return; 723 724 out_unlock: 725 up_read(&ufile->hw_destroy_rwsem); 726 out_zap: 727 /* 728 * We can't allow the VMA to be created with the actual IO pages, that 729 * would break our API contract, and it can't be stopped at this 730 * point, so zap it. 731 */ 732 vma->vm_private_data = NULL; 733 zap_vma_ptes(vma, vma->vm_start, vma->vm_end - vma->vm_start); 734 } 735 736 static void rdma_umap_close(struct vm_area_struct *vma) 737 { 738 struct ib_uverbs_file *ufile = vma->vm_file->private_data; 739 struct rdma_umap_priv *priv = vma->vm_private_data; 740 741 if (!priv) 742 return; 743 744 /* 745 * The vma holds a reference on the struct file that created it, which 746 * in turn means that the ib_uverbs_file is guaranteed to exist at 747 * this point. 748 */ 749 mutex_lock(&ufile->umap_lock); 750 751 list_del(&priv->list); 752 mutex_unlock(&ufile->umap_lock); 753 kfree(priv); 754 } 755 756 static const struct vm_operations_struct rdma_umap_ops = { 757 .open = rdma_umap_open, 758 .close = rdma_umap_close, 759 }; 760 761 void uverbs_user_mmap_disassociate(struct ib_uverbs_file *ufile) 762 { 763 struct rdma_umap_priv *priv, *next_priv; 764 765 lockdep_assert_held(&ufile->hw_destroy_rwsem); 766 767 while (1) { 768 struct mm_struct *mm = NULL; 769 770 /* Get an arbitrary mm pointer that hasn't been cleaned yet */ 771 mutex_lock(&ufile->umap_lock); 772 while (!list_empty(&ufile->umaps)) { 773 int ret; 774 775 priv = list_first_entry(&ufile->umaps, 776 struct rdma_umap_priv, list); 777 mm = priv->vma->vm_mm; 778 ret = mmget_not_zero(mm); 779 if (!ret) { 780 list_del_init(&priv->list); 781 mm = NULL; 782 continue; 783 } 784 break; 785 } 786 mutex_unlock(&ufile->umap_lock); 787 if (!mm) 788 return; 789 790 /* 791 * The umap_lock is nested under mmap_sem since it used within 792 * the vma_ops callbacks, so we have to clean the list one mm 793 * at a time to get the lock ordering right. Typically there 794 * will only be one mm, so no big deal. 795 */ 796 down_read(&mm->mmap_sem); 797 mutex_lock(&ufile->umap_lock); 798 list_for_each_entry_safe (priv, next_priv, &ufile->umaps, 799 list) { 800 struct vm_area_struct *vma = priv->vma; 801 802 if (vma->vm_mm != mm) 803 continue; 804 list_del_init(&priv->list); 805 806 zap_vma_ptes(vma, vma->vm_start, 807 vma->vm_end - vma->vm_start); 808 } 809 mutex_unlock(&ufile->umap_lock); 810 up_read(&mm->mmap_sem); 811 mmput(mm); 812 } 813 } 814 815 /* 816 * ib_uverbs_open() does not need the BKL: 817 * 818 * - the ib_uverbs_device structures are properly reference counted and 819 * everything else is purely local to the file being created, so 820 * races against other open calls are not a problem; 821 * - there is no ioctl method to race against; 822 * - the open method will either immediately run -ENXIO, or all 823 * required initialization will be done. 824 */ 825 static int ib_uverbs_open(struct inode *inode, struct file *filp) 826 { 827 struct ib_uverbs_device *dev; 828 struct ib_uverbs_file *file; 829 struct ib_device *ib_dev; 830 int ret; 831 int module_dependent; 832 int srcu_key; 833 834 dev = container_of(inode->i_cdev->si_drv1, struct ib_uverbs_device, cdev); 835 if (!atomic_inc_not_zero(&dev->refcount)) 836 return -ENXIO; 837 838 get_device(&dev->dev); 839 srcu_key = srcu_read_lock(&dev->disassociate_srcu); 840 mutex_lock(&dev->lists_mutex); 841 ib_dev = srcu_dereference(dev->ib_dev, 842 &dev->disassociate_srcu); 843 if (!ib_dev) { 844 ret = -EIO; 845 goto err; 846 } 847 848 /* In case IB device supports disassociate ucontext, there is no hard 849 * dependency between uverbs device and its low level device. 850 */ 851 module_dependent = !(ib_dev->disassociate_ucontext); 852 853 if (module_dependent) { 854 if (!try_module_get(ib_dev->owner)) { 855 ret = -ENODEV; 856 goto err; 857 } 858 } 859 860 file = kzalloc(sizeof(*file), GFP_KERNEL); 861 if (!file) { 862 ret = -ENOMEM; 863 if (module_dependent) 864 goto err_module; 865 866 goto err; 867 } 868 869 file->device = dev; 870 kref_init(&file->ref); 871 mutex_init(&file->ucontext_lock); 872 873 spin_lock_init(&file->uobjects_lock); 874 INIT_LIST_HEAD(&file->uobjects); 875 init_rwsem(&file->hw_destroy_rwsem); 876 mutex_init(&file->umap_lock); 877 INIT_LIST_HEAD(&file->umaps); 878 879 filp->private_data = file; 880 list_add_tail(&file->list, &dev->uverbs_file_list); 881 mutex_unlock(&dev->lists_mutex); 882 srcu_read_unlock(&dev->disassociate_srcu, srcu_key); 883 884 setup_ufile_idr_uobject(file); 885 886 return nonseekable_open(inode, filp); 887 888 err_module: 889 module_put(ib_dev->owner); 890 891 err: 892 mutex_unlock(&dev->lists_mutex); 893 srcu_read_unlock(&dev->disassociate_srcu, srcu_key); 894 if (atomic_dec_and_test(&dev->refcount)) 895 ib_uverbs_comp_dev(dev); 896 897 put_device(&dev->dev); 898 return ret; 899 } 900 901 static int ib_uverbs_close(struct inode *inode, struct file *filp) 902 { 903 struct ib_uverbs_file *file = filp->private_data; 904 905 uverbs_destroy_ufile_hw(file, RDMA_REMOVE_CLOSE); 906 907 mutex_lock(&file->device->lists_mutex); 908 list_del_init(&file->list); 909 mutex_unlock(&file->device->lists_mutex); 910 911 kref_put(&file->ref, ib_uverbs_release_file); 912 913 return 0; 914 } 915 916 static const struct file_operations uverbs_fops = { 917 .owner = THIS_MODULE, 918 .write = ib_uverbs_write, 919 .open = ib_uverbs_open, 920 .release = ib_uverbs_close, 921 .llseek = no_llseek, 922 .unlocked_ioctl = ib_uverbs_ioctl, 923 .compat_ioctl = NULL, 924 }; 925 926 static const struct file_operations uverbs_mmap_fops = { 927 .owner = THIS_MODULE, 928 .write = ib_uverbs_write, 929 .mmap = ib_uverbs_mmap, 930 .open = ib_uverbs_open, 931 .release = ib_uverbs_close, 932 .llseek = no_llseek, 933 .unlocked_ioctl = ib_uverbs_ioctl, 934 .compat_ioctl = NULL, 935 }; 936 937 static struct ib_client uverbs_client = { 938 .name = "uverbs", 939 .add = ib_uverbs_add_one, 940 .remove = ib_uverbs_remove_one 941 }; 942 943 static ssize_t ibdev_show(struct device *device, struct device_attribute *attr, 944 char *buf) 945 { 946 struct ib_uverbs_device *dev = 947 container_of(device, struct ib_uverbs_device, dev); 948 int ret = -ENODEV; 949 int srcu_key; 950 struct ib_device *ib_dev; 951 952 srcu_key = srcu_read_lock(&dev->disassociate_srcu); 953 ib_dev = srcu_dereference(dev->ib_dev, &dev->disassociate_srcu); 954 if (ib_dev) 955 ret = sprintf(buf, "%s\n", ib_dev->name); 956 srcu_read_unlock(&dev->disassociate_srcu, srcu_key); 957 958 return ret; 959 } 960 static DEVICE_ATTR(ibdev, S_IRUGO, ibdev_show, NULL); 961 962 static ssize_t abi_version_show(struct device *device, 963 struct device_attribute *attr, char *buf) 964 { 965 struct ib_uverbs_device *dev = 966 container_of(device, struct ib_uverbs_device, dev); 967 int ret = -ENODEV; 968 int srcu_key; 969 struct ib_device *ib_dev; 970 971 srcu_key = srcu_read_lock(&dev->disassociate_srcu); 972 ib_dev = srcu_dereference(dev->ib_dev, &dev->disassociate_srcu); 973 if (ib_dev) 974 ret = sprintf(buf, "%d\n", ib_dev->uverbs_abi_ver); 975 srcu_read_unlock(&dev->disassociate_srcu, srcu_key); 976 977 return ret; 978 } 979 static DEVICE_ATTR(abi_version, S_IRUGO, abi_version_show, NULL); 980 981 static CLASS_ATTR_STRING(abi_version, S_IRUGO, 982 __stringify(IB_USER_VERBS_ABI_VERSION)); 983 984 static int ib_uverbs_create_uapi(struct ib_device *device, 985 struct ib_uverbs_device *uverbs_dev) 986 { 987 struct uverbs_api *uapi; 988 989 uapi = uverbs_alloc_api(device); 990 if (IS_ERR(uapi)) 991 return PTR_ERR(uapi); 992 993 uverbs_dev->uapi = uapi; 994 return 0; 995 } 996 997 static ssize_t 998 show_dev_device(struct device *device, struct device_attribute *attr, char *buf) 999 { 1000 struct ib_uverbs_device *dev = 1001 container_of(device, struct ib_uverbs_device, dev); 1002 1003 if (!dev->ib_dev->dma_device) 1004 return -ENODEV; 1005 1006 return sprintf(buf, "0x%04x\n", 1007 ((struct pci_dev *)dev->ib_dev->dma_device)->device); 1008 } 1009 static DEVICE_ATTR(device, S_IRUGO, show_dev_device, NULL); 1010 1011 static ssize_t 1012 show_dev_vendor(struct device *device, struct device_attribute *attr, char *buf) 1013 { 1014 struct ib_uverbs_device *dev = 1015 container_of(device, struct ib_uverbs_device, dev); 1016 1017 if (!dev->ib_dev->dma_device) 1018 return -ENODEV; 1019 1020 return sprintf(buf, "0x%04x\n", 1021 ((struct pci_dev *)dev->ib_dev->dma_device)->vendor); 1022 } 1023 static DEVICE_ATTR(vendor, S_IRUGO, show_dev_vendor, NULL); 1024 1025 static struct attribute *ib_dev_attrs[] = { 1026 &dev_attr_device.attr, 1027 &dev_attr_vendor.attr, 1028 NULL, 1029 }; 1030 1031 static const struct attribute_group dev_attr_group = { 1032 .name = "device", 1033 .attrs = ib_dev_attrs, 1034 }; 1035 1036 static void ib_uverbs_add_one(struct ib_device *device) 1037 { 1038 int devnum; 1039 dev_t base; 1040 struct ib_uverbs_device *uverbs_dev; 1041 int ret; 1042 1043 if (!device->alloc_ucontext) 1044 return; 1045 1046 uverbs_dev = kzalloc(sizeof(*uverbs_dev), GFP_KERNEL); 1047 if (!uverbs_dev) 1048 return; 1049 1050 ret = init_srcu_struct(&uverbs_dev->disassociate_srcu); 1051 if (ret) { 1052 kfree(uverbs_dev); 1053 return; 1054 } 1055 1056 uverbs_dev->dev.class = uverbs_class; 1057 uverbs_dev->dev.parent = device->dev.parent; 1058 uverbs_dev->dev.release = ib_uverbs_release_dev; 1059 device_initialize(&uverbs_dev->dev); 1060 atomic_set(&uverbs_dev->refcount, 1); 1061 init_completion(&uverbs_dev->comp); 1062 uverbs_dev->xrcd_tree = RB_ROOT; 1063 mutex_init(&uverbs_dev->xrcd_tree_mutex); 1064 mutex_init(&uverbs_dev->lists_mutex); 1065 INIT_LIST_HEAD(&uverbs_dev->uverbs_file_list); 1066 rcu_assign_pointer(uverbs_dev->ib_dev, device); 1067 uverbs_dev->num_comp_vectors = device->num_comp_vectors; 1068 1069 devnum = ida_alloc_max(&uverbs_ida, IB_UVERBS_MAX_DEVICES - 1, 1070 GFP_KERNEL); 1071 if (devnum < 0) 1072 goto err; 1073 uverbs_dev->devnum = devnum; 1074 if (devnum >= IB_UVERBS_NUM_FIXED_MINOR) 1075 base = dynamic_uverbs_dev + devnum - IB_UVERBS_NUM_FIXED_MINOR; 1076 else 1077 base = IB_UVERBS_BASE_DEV + devnum; 1078 1079 if (ib_uverbs_create_uapi(device, uverbs_dev)) 1080 goto err_uapi; 1081 1082 uverbs_dev->dev.devt = base; 1083 dev_set_name(&uverbs_dev->dev, "uverbs%d", uverbs_dev->devnum); 1084 1085 cdev_init(&uverbs_dev->cdev, 1086 device->mmap ? &uverbs_mmap_fops : &uverbs_fops); 1087 uverbs_dev->cdev.owner = THIS_MODULE; 1088 1089 kobject_set_name(&uverbs_dev->cdev.kobj, "uverbs%d", uverbs_dev->devnum); 1090 ret = cdev_device_add(&uverbs_dev->cdev, &uverbs_dev->dev); 1091 if (ret) 1092 goto err_uapi; 1093 1094 if (device_create_file(&uverbs_dev->dev, &dev_attr_ibdev)) 1095 goto err_cdev; 1096 if (device_create_file(&uverbs_dev->dev, &dev_attr_abi_version)) 1097 goto err_cdev; 1098 if (sysfs_create_group(&uverbs_dev->dev.kobj, &dev_attr_group)) 1099 goto err_cdev; 1100 1101 ib_set_client_data(device, &uverbs_client, uverbs_dev); 1102 return; 1103 1104 err_cdev: 1105 cdev_device_del(&uverbs_dev->cdev, &uverbs_dev->dev); 1106 err_uapi: 1107 ida_free(&uverbs_ida, devnum); 1108 err: 1109 if (atomic_dec_and_test(&uverbs_dev->refcount)) 1110 ib_uverbs_comp_dev(uverbs_dev); 1111 wait_for_completion(&uverbs_dev->comp); 1112 put_device(&uverbs_dev->dev); 1113 return; 1114 } 1115 1116 static void ib_uverbs_free_hw_resources(struct ib_uverbs_device *uverbs_dev, 1117 struct ib_device *ib_dev) 1118 { 1119 struct ib_uverbs_file *file; 1120 1121 /* Pending running commands to terminate */ 1122 uverbs_disassociate_api_pre(uverbs_dev); 1123 1124 mutex_lock(&uverbs_dev->lists_mutex); 1125 while (!list_empty(&uverbs_dev->uverbs_file_list)) { 1126 file = list_first_entry(&uverbs_dev->uverbs_file_list, 1127 struct ib_uverbs_file, list); 1128 list_del_init(&file->list); 1129 kref_get(&file->ref); 1130 1131 /* We must release the mutex before going ahead and calling 1132 * uverbs_cleanup_ufile, as it might end up indirectly calling 1133 * uverbs_close, for example due to freeing the resources (e.g 1134 * mmput). 1135 */ 1136 mutex_unlock(&uverbs_dev->lists_mutex); 1137 1138 ib_uverbs_async_handler(READ_ONCE(file->async_file), 0, 1139 IB_EVENT_DEVICE_FATAL, NULL, NULL); 1140 1141 uverbs_destroy_ufile_hw(file, RDMA_REMOVE_DRIVER_REMOVE); 1142 kref_put(&file->ref, ib_uverbs_release_file); 1143 1144 mutex_lock(&uverbs_dev->lists_mutex); 1145 } 1146 mutex_unlock(&uverbs_dev->lists_mutex); 1147 1148 uverbs_disassociate_api(uverbs_dev->uapi); 1149 } 1150 1151 static void ib_uverbs_remove_one(struct ib_device *device, void *client_data) 1152 { 1153 struct ib_uverbs_device *uverbs_dev = client_data; 1154 int wait_clients = 1; 1155 1156 if (!uverbs_dev) 1157 return; 1158 1159 cdev_device_del(&uverbs_dev->cdev, &uverbs_dev->dev); 1160 ida_free(&uverbs_ida, uverbs_dev->devnum); 1161 1162 if (device->disassociate_ucontext) { 1163 /* We disassociate HW resources and immediately return. 1164 * Userspace will see a EIO errno for all future access. 1165 * Upon returning, ib_device may be freed internally and is not 1166 * valid any more. 1167 * uverbs_device is still available until all clients close 1168 * their files, then the uverbs device ref count will be zero 1169 * and its resources will be freed. 1170 * Note: At this point no more files can be opened since the 1171 * cdev was deleted, however active clients can still issue 1172 * commands and close their open files. 1173 */ 1174 ib_uverbs_free_hw_resources(uverbs_dev, device); 1175 wait_clients = 0; 1176 } 1177 1178 if (atomic_dec_and_test(&uverbs_dev->refcount)) 1179 ib_uverbs_comp_dev(uverbs_dev); 1180 if (wait_clients) 1181 wait_for_completion(&uverbs_dev->comp); 1182 1183 put_device(&uverbs_dev->dev); 1184 } 1185 1186 static char *uverbs_devnode(struct device *dev, umode_t *mode) 1187 { 1188 if (mode) 1189 *mode = 0666; 1190 return kasprintf(GFP_KERNEL, "infiniband/%s", dev_name(dev)); 1191 } 1192 1193 static int __init ib_uverbs_init(void) 1194 { 1195 int ret; 1196 1197 ret = register_chrdev_region(IB_UVERBS_BASE_DEV, 1198 IB_UVERBS_NUM_FIXED_MINOR, 1199 "infiniband_verbs"); 1200 if (ret) { 1201 pr_err("user_verbs: couldn't register device number\n"); 1202 goto out; 1203 } 1204 1205 ret = alloc_chrdev_region(&dynamic_uverbs_dev, 0, 1206 IB_UVERBS_NUM_DYNAMIC_MINOR, 1207 "infiniband_verbs"); 1208 if (ret) { 1209 pr_err("couldn't register dynamic device number\n"); 1210 goto out_alloc; 1211 } 1212 1213 uverbs_class = class_create(THIS_MODULE, "infiniband_verbs"); 1214 if (IS_ERR(uverbs_class)) { 1215 ret = PTR_ERR(uverbs_class); 1216 pr_err("user_verbs: couldn't create class infiniband_verbs\n"); 1217 goto out_chrdev; 1218 } 1219 1220 uverbs_class->devnode = uverbs_devnode; 1221 1222 ret = class_create_file(uverbs_class, &class_attr_abi_version.attr); 1223 if (ret) { 1224 pr_err("user_verbs: couldn't create abi_version attribute\n"); 1225 goto out_class; 1226 } 1227 1228 ret = ib_register_client(&uverbs_client); 1229 if (ret) { 1230 pr_err("user_verbs: couldn't register client\n"); 1231 goto out_class; 1232 } 1233 1234 return 0; 1235 1236 out_class: 1237 class_destroy(uverbs_class); 1238 1239 out_chrdev: 1240 unregister_chrdev_region(dynamic_uverbs_dev, 1241 IB_UVERBS_NUM_DYNAMIC_MINOR); 1242 1243 out_alloc: 1244 unregister_chrdev_region(IB_UVERBS_BASE_DEV, 1245 IB_UVERBS_NUM_FIXED_MINOR); 1246 1247 out: 1248 return ret; 1249 } 1250 1251 static void __exit ib_uverbs_cleanup(void) 1252 { 1253 ib_unregister_client(&uverbs_client); 1254 class_destroy(uverbs_class); 1255 unregister_chrdev_region(IB_UVERBS_BASE_DEV, 1256 IB_UVERBS_NUM_FIXED_MINOR); 1257 unregister_chrdev_region(dynamic_uverbs_dev, 1258 IB_UVERBS_NUM_DYNAMIC_MINOR); 1259 } 1260 1261 module_init_order(ib_uverbs_init, SI_ORDER_FIFTH); 1262 module_exit_order(ib_uverbs_cleanup, SI_ORDER_FIFTH); 1263