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 #include <linux/module.h> 41 #include <linux/device.h> 42 #include <linux/err.h> 43 #include <linux/fs.h> 44 #include <linux/poll.h> 45 #include <linux/sched.h> 46 #include <linux/file.h> 47 #include <linux/cdev.h> 48 #include <linux/slab.h> 49 #include <linux/pci.h> 50 #include <linux/lockdep.h> 51 52 #include <asm/uaccess.h> 53 54 #include <rdma/ib.h> 55 #include <rdma/uverbs_std_types.h> 56 57 #include "uverbs.h" 58 #include "core_priv.h" 59 #include "rdma_core.h" 60 61 MODULE_AUTHOR("Roland Dreier"); 62 MODULE_DESCRIPTION("InfiniBand userspace verbs access"); 63 MODULE_LICENSE("Dual BSD/GPL"); 64 65 enum { 66 IB_UVERBS_MAJOR = 231, 67 IB_UVERBS_BASE_MINOR = 192, 68 IB_UVERBS_MAX_DEVICES = RDMA_MAX_PORTS, 69 IB_UVERBS_NUM_FIXED_MINOR = 32, 70 IB_UVERBS_NUM_DYNAMIC_MINOR = IB_UVERBS_MAX_DEVICES - IB_UVERBS_NUM_FIXED_MINOR, 71 }; 72 73 #define IB_UVERBS_BASE_DEV MKDEV(IB_UVERBS_MAJOR, IB_UVERBS_BASE_MINOR) 74 75 static dev_t dynamic_uverbs_dev; 76 static struct class *uverbs_class; 77 78 static DEFINE_IDA(uverbs_ida); 79 static void ib_uverbs_add_one(struct ib_device *device); 80 static void ib_uverbs_remove_one(struct ib_device *device, void *client_data); 81 82 /* 83 * Must be called with the ufile->device->disassociate_srcu held, and the lock 84 * must be held until use of the ucontext is finished. 85 */ 86 struct ib_ucontext *ib_uverbs_get_ucontext_file(struct ib_uverbs_file *ufile) 87 { 88 /* 89 * We do not hold the hw_destroy_rwsem lock for this flow, instead 90 * srcu is used. It does not matter if someone races this with 91 * get_context, we get NULL or valid ucontext. 92 */ 93 struct ib_ucontext *ucontext = READ_ONCE(ufile->ucontext); 94 95 if (!srcu_dereference(ufile->device->ib_dev, 96 &ufile->device->disassociate_srcu)) 97 return ERR_PTR(-EIO); 98 99 if (!ucontext) 100 return ERR_PTR(-EINVAL); 101 102 return ucontext; 103 } 104 EXPORT_SYMBOL(ib_uverbs_get_ucontext_file); 105 106 int uverbs_dealloc_mw(struct ib_mw *mw) 107 { 108 struct ib_pd *pd = mw->pd; 109 int ret; 110 111 ret = mw->device->dealloc_mw(mw); 112 if (!ret) 113 atomic_dec(&pd->usecnt); 114 return ret; 115 } 116 117 static void ib_uverbs_release_dev(struct device *device) 118 { 119 struct ib_uverbs_device *dev = 120 container_of(device, struct ib_uverbs_device, dev); 121 122 uverbs_destroy_api(dev->uapi); 123 cleanup_srcu_struct(&dev->disassociate_srcu); 124 mutex_destroy(&dev->lists_mutex); 125 mutex_destroy(&dev->xrcd_tree_mutex); 126 kfree(dev); 127 } 128 129 void ib_uverbs_release_ucq(struct ib_uverbs_completion_event_file *ev_file, 130 struct ib_ucq_object *uobj) 131 { 132 struct ib_uverbs_event *evt, *tmp; 133 134 if (ev_file) { 135 spin_lock_irq(&ev_file->ev_queue.lock); 136 list_for_each_entry_safe(evt, tmp, &uobj->comp_list, obj_list) { 137 list_del(&evt->list); 138 kfree(evt); 139 } 140 spin_unlock_irq(&ev_file->ev_queue.lock); 141 142 uverbs_uobject_put(&ev_file->uobj); 143 } 144 145 ib_uverbs_release_uevent(&uobj->uevent); 146 } 147 148 void ib_uverbs_release_uevent(struct ib_uevent_object *uobj) 149 { 150 struct ib_uverbs_async_event_file *async_file = 151 READ_ONCE(uobj->uobject.ufile->async_file); 152 struct ib_uverbs_event *evt, *tmp; 153 154 if (!async_file) 155 return; 156 157 spin_lock_irq(&async_file->ev_queue.lock); 158 list_for_each_entry_safe(evt, tmp, &uobj->event_list, obj_list) { 159 list_del(&evt->list); 160 kfree(evt); 161 } 162 spin_unlock_irq(&async_file->ev_queue.lock); 163 } 164 165 void ib_uverbs_detach_umcast(struct ib_qp *qp, 166 struct ib_uqp_object *uobj) 167 { 168 struct ib_uverbs_mcast_entry *mcast, *tmp; 169 170 list_for_each_entry_safe(mcast, tmp, &uobj->mcast_list, list) { 171 ib_detach_mcast(qp, &mcast->gid, mcast->lid); 172 list_del(&mcast->list); 173 kfree(mcast); 174 } 175 } 176 177 static void ib_uverbs_comp_dev(struct ib_uverbs_device *dev) 178 { 179 complete(&dev->comp); 180 } 181 182 void ib_uverbs_release_file(struct kref *ref) 183 { 184 struct ib_uverbs_file *file = 185 container_of(ref, struct ib_uverbs_file, ref); 186 struct ib_device *ib_dev; 187 int srcu_key; 188 189 release_ufile_idr_uobject(file); 190 191 srcu_key = srcu_read_lock(&file->device->disassociate_srcu); 192 ib_dev = srcu_dereference(file->device->ib_dev, 193 &file->device->disassociate_srcu); 194 if (ib_dev && !ib_dev->disassociate_ucontext) 195 module_put(ib_dev->owner); 196 srcu_read_unlock(&file->device->disassociate_srcu, srcu_key); 197 198 if (atomic_dec_and_test(&file->device->refcount)) 199 ib_uverbs_comp_dev(file->device); 200 201 if (file->async_file) 202 uverbs_uobject_put(&file->async_file->uobj); 203 put_device(&file->device->dev); 204 205 mutex_destroy(&file->umap_lock); 206 mutex_destroy(&file->ucontext_lock); 207 kfree(file); 208 } 209 210 static ssize_t ib_uverbs_event_read(struct ib_uverbs_event_queue *ev_queue, 211 struct file *filp, char __user *buf, 212 size_t count, loff_t *pos, 213 size_t eventsz) 214 { 215 struct ib_uverbs_event *event; 216 int ret = 0; 217 218 spin_lock_irq(&ev_queue->lock); 219 220 while (list_empty(&ev_queue->event_list)) { 221 spin_unlock_irq(&ev_queue->lock); 222 223 if (filp->f_flags & O_NONBLOCK) 224 return -EAGAIN; 225 226 if (wait_event_interruptible(ev_queue->poll_wait, 227 (!list_empty(&ev_queue->event_list) || 228 ev_queue->is_closed))) 229 return -ERESTARTSYS; 230 231 spin_lock_irq(&ev_queue->lock); 232 233 /* If device was disassociated and no event exists set an error */ 234 if (list_empty(&ev_queue->event_list) && ev_queue->is_closed) { 235 spin_unlock_irq(&ev_queue->lock); 236 return -EIO; 237 } 238 } 239 240 event = list_entry(ev_queue->event_list.next, struct ib_uverbs_event, list); 241 242 if (eventsz > count) { 243 ret = -EINVAL; 244 event = NULL; 245 } else { 246 list_del(ev_queue->event_list.next); 247 if (event->counter) { 248 ++(*event->counter); 249 list_del(&event->obj_list); 250 } 251 } 252 253 spin_unlock_irq(&ev_queue->lock); 254 255 if (event) { 256 if (copy_to_user(buf, event, eventsz)) 257 ret = -EFAULT; 258 else 259 ret = eventsz; 260 } 261 262 kfree(event); 263 264 return ret; 265 } 266 267 static ssize_t ib_uverbs_async_event_read(struct file *filp, char __user *buf, 268 size_t count, loff_t *pos) 269 { 270 struct ib_uverbs_async_event_file *file = filp->private_data; 271 272 return ib_uverbs_event_read(&file->ev_queue, filp, buf, count, pos, 273 sizeof(struct ib_uverbs_async_event_desc)); 274 } 275 276 static ssize_t ib_uverbs_comp_event_read(struct file *filp, char __user *buf, 277 size_t count, loff_t *pos) 278 { 279 struct ib_uverbs_completion_event_file *comp_ev_file = 280 filp->private_data; 281 282 return ib_uverbs_event_read(&comp_ev_file->ev_queue, filp, buf, count, 283 pos, 284 sizeof(struct ib_uverbs_comp_event_desc)); 285 } 286 287 static __poll_t ib_uverbs_event_poll(struct ib_uverbs_event_queue *ev_queue, 288 struct file *filp, 289 struct poll_table_struct *wait) 290 { 291 __poll_t pollflags = 0; 292 293 poll_wait(filp, &ev_queue->poll_wait, wait); 294 295 spin_lock_irq(&ev_queue->lock); 296 if (!list_empty(&ev_queue->event_list)) 297 pollflags = POLLIN | POLLRDNORM; 298 spin_unlock_irq(&ev_queue->lock); 299 300 return pollflags; 301 } 302 303 static __poll_t ib_uverbs_async_event_poll(struct file *filp, 304 struct poll_table_struct *wait) 305 { 306 struct ib_uverbs_async_event_file *file = filp->private_data; 307 308 return ib_uverbs_event_poll(&file->ev_queue, filp, wait); 309 } 310 311 static __poll_t ib_uverbs_comp_event_poll(struct file *filp, 312 struct poll_table_struct *wait) 313 { 314 struct ib_uverbs_completion_event_file *comp_ev_file = 315 filp->private_data; 316 317 return ib_uverbs_event_poll(&comp_ev_file->ev_queue, filp, wait); 318 } 319 320 static int ib_uverbs_async_event_fasync(int fd, struct file *filp, int on) 321 { 322 struct ib_uverbs_async_event_file *file = filp->private_data; 323 324 return fasync_helper(fd, filp, on, &file->ev_queue.async_queue); 325 } 326 327 static int ib_uverbs_comp_event_fasync(int fd, struct file *filp, int on) 328 { 329 struct ib_uverbs_completion_event_file *comp_ev_file = 330 filp->private_data; 331 332 return fasync_helper(fd, filp, on, &comp_ev_file->ev_queue.async_queue); 333 } 334 335 const struct file_operations uverbs_event_fops = { 336 .owner = THIS_MODULE, 337 .read = ib_uverbs_comp_event_read, 338 .poll = ib_uverbs_comp_event_poll, 339 .release = uverbs_uobject_fd_release, 340 .fasync = ib_uverbs_comp_event_fasync, 341 .llseek = no_llseek, 342 }; 343 344 const struct file_operations uverbs_async_event_fops = { 345 .owner = THIS_MODULE, 346 .read = ib_uverbs_async_event_read, 347 .poll = ib_uverbs_async_event_poll, 348 .release = uverbs_uobject_fd_release, 349 .fasync = ib_uverbs_async_event_fasync, 350 .llseek = no_llseek, 351 }; 352 353 void ib_uverbs_comp_handler(struct ib_cq *cq, void *cq_context) 354 { 355 struct ib_uverbs_event_queue *ev_queue = cq_context; 356 struct ib_ucq_object *uobj; 357 struct ib_uverbs_event *entry; 358 unsigned long flags; 359 360 if (!ev_queue) 361 return; 362 363 spin_lock_irqsave(&ev_queue->lock, flags); 364 if (ev_queue->is_closed) { 365 spin_unlock_irqrestore(&ev_queue->lock, flags); 366 return; 367 } 368 369 entry = kmalloc(sizeof(*entry), GFP_ATOMIC); 370 if (!entry) { 371 spin_unlock_irqrestore(&ev_queue->lock, flags); 372 return; 373 } 374 375 uobj = cq->uobject; 376 377 entry->desc.comp.cq_handle = cq->uobject->uevent.uobject.user_handle; 378 entry->counter = &uobj->comp_events_reported; 379 380 list_add_tail(&entry->list, &ev_queue->event_list); 381 list_add_tail(&entry->obj_list, &uobj->comp_list); 382 spin_unlock_irqrestore(&ev_queue->lock, flags); 383 384 wake_up_interruptible(&ev_queue->poll_wait); 385 kill_fasync(&ev_queue->async_queue, SIGIO, POLL_IN); 386 } 387 388 static void 389 ib_uverbs_async_handler(struct ib_uverbs_async_event_file *async_file, 390 __u64 element, __u64 event, struct list_head *obj_list, 391 u32 *counter) 392 { 393 struct ib_uverbs_event *entry; 394 unsigned long flags; 395 396 if (!async_file) 397 return; 398 399 spin_lock_irqsave(&async_file->ev_queue.lock, flags); 400 if (async_file->ev_queue.is_closed) { 401 spin_unlock_irqrestore(&async_file->ev_queue.lock, flags); 402 return; 403 } 404 405 entry = kmalloc(sizeof(*entry), GFP_ATOMIC); 406 if (!entry) { 407 spin_unlock_irqrestore(&async_file->ev_queue.lock, flags); 408 return; 409 } 410 411 entry->desc.async.element = element; 412 entry->desc.async.event_type = event; 413 entry->desc.async.reserved = 0; 414 entry->counter = counter; 415 416 list_add_tail(&entry->list, &async_file->ev_queue.event_list); 417 if (obj_list) 418 list_add_tail(&entry->obj_list, obj_list); 419 spin_unlock_irqrestore(&async_file->ev_queue.lock, flags); 420 421 wake_up_interruptible(&async_file->ev_queue.poll_wait); 422 kill_fasync(&async_file->ev_queue.async_queue, SIGIO, POLL_IN); 423 } 424 425 static void uverbs_uobj_event(struct ib_uevent_object *eobj, 426 struct ib_event *event) 427 { 428 ib_uverbs_async_handler(READ_ONCE(eobj->uobject.ufile->async_file), 429 eobj->uobject.user_handle, event->event, 430 &eobj->event_list, &eobj->events_reported); 431 } 432 433 void ib_uverbs_cq_event_handler(struct ib_event *event, void *context_ptr) 434 { 435 uverbs_uobj_event(&event->element.cq->uobject->uevent, event); 436 } 437 438 void ib_uverbs_qp_event_handler(struct ib_event *event, void *context_ptr) 439 { 440 /* for XRC target qp's, check that qp is live */ 441 if (!event->element.qp->uobject) 442 return; 443 444 uverbs_uobj_event(&event->element.qp->uobject->uevent, event); 445 } 446 447 void ib_uverbs_wq_event_handler(struct ib_event *event, void *context_ptr) 448 { 449 uverbs_uobj_event(&event->element.wq->uobject->uevent, event); 450 } 451 452 void ib_uverbs_srq_event_handler(struct ib_event *event, void *context_ptr) 453 { 454 uverbs_uobj_event(&event->element.srq->uobject->uevent, event); 455 } 456 457 static void ib_uverbs_event_handler(struct ib_event_handler *handler, 458 struct ib_event *event) 459 { 460 ib_uverbs_async_handler( 461 container_of(handler, struct ib_uverbs_async_event_file, 462 event_handler), 463 event->element.port_num, event->event, NULL, NULL); 464 } 465 466 void ib_uverbs_init_event_queue(struct ib_uverbs_event_queue *ev_queue) 467 { 468 spin_lock_init(&ev_queue->lock); 469 INIT_LIST_HEAD(&ev_queue->event_list); 470 init_waitqueue_head(&ev_queue->poll_wait); 471 ev_queue->is_closed = 0; 472 ev_queue->async_queue = NULL; 473 } 474 475 void ib_uverbs_init_async_event_file( 476 struct ib_uverbs_async_event_file *async_file) 477 { 478 struct ib_uverbs_file *uverbs_file = async_file->uobj.ufile; 479 struct ib_device *ib_dev = async_file->uobj.context->device; 480 481 ib_uverbs_init_event_queue(&async_file->ev_queue); 482 483 /* The first async_event_file becomes the default one for the file. */ 484 mutex_lock(&uverbs_file->ucontext_lock); 485 if (!uverbs_file->async_file) { 486 /* Pairs with the put in ib_uverbs_release_file */ 487 uverbs_uobject_get(&async_file->uobj); 488 atomic_store_rel_ptr((uintptr_t *)&uverbs_file->async_file, (uintptr_t)async_file); 489 } 490 mutex_unlock(&uverbs_file->ucontext_lock); 491 492 INIT_IB_EVENT_HANDLER(&async_file->event_handler, ib_dev, 493 ib_uverbs_event_handler); 494 ib_register_event_handler(&async_file->event_handler); 495 } 496 497 static ssize_t verify_hdr(struct ib_uverbs_cmd_hdr *hdr, 498 struct ib_uverbs_ex_cmd_hdr *ex_hdr, size_t count, 499 const struct uverbs_api_write_method *method_elm) 500 { 501 if (method_elm->is_ex) { 502 count -= sizeof(*hdr) + sizeof(*ex_hdr); 503 504 if ((hdr->in_words + ex_hdr->provider_in_words) * 8 != count) 505 return -EINVAL; 506 507 if (hdr->in_words * 8 < method_elm->req_size) 508 return -ENOSPC; 509 510 if (ex_hdr->cmd_hdr_reserved) 511 return -EINVAL; 512 513 if (ex_hdr->response) { 514 if (!hdr->out_words && !ex_hdr->provider_out_words) 515 return -EINVAL; 516 517 if (hdr->out_words * 8 < method_elm->resp_size) 518 return -ENOSPC; 519 520 if (!access_ok(u64_to_user_ptr(ex_hdr->response), 521 (hdr->out_words + ex_hdr->provider_out_words) * 8)) 522 return -EFAULT; 523 } else { 524 if (hdr->out_words || ex_hdr->provider_out_words) 525 return -EINVAL; 526 } 527 528 return 0; 529 } 530 531 /* not extended command */ 532 if (hdr->in_words * 4 != count) 533 return -EINVAL; 534 535 if (count < method_elm->req_size + sizeof(hdr)) { 536 /* 537 * rdma-core v18 and v19 have a bug where they send DESTROY_CQ 538 * with a 16 byte write instead of 24. Old kernels didn't 539 * check the size so they allowed this. Now that the size is 540 * checked provide a compatibility work around to not break 541 * those userspaces. 542 */ 543 if (hdr->command == IB_USER_VERBS_CMD_DESTROY_CQ && 544 count == 16) { 545 hdr->in_words = 6; 546 return 0; 547 } 548 return -ENOSPC; 549 } 550 if (hdr->out_words * 4 < method_elm->resp_size) 551 return -ENOSPC; 552 553 return 0; 554 } 555 556 static ssize_t ib_uverbs_write(struct file *filp, const char __user *buf, 557 size_t count, loff_t *pos) 558 { 559 struct ib_uverbs_file *file = filp->private_data; 560 const struct uverbs_api_write_method *method_elm; 561 struct uverbs_api *uapi = file->device->uapi; 562 struct ib_uverbs_ex_cmd_hdr ex_hdr; 563 struct ib_uverbs_cmd_hdr hdr; 564 struct uverbs_attr_bundle bundle; 565 int srcu_key; 566 ssize_t ret; 567 568 if (!ib_safe_file_access(filp)) { 569 pr_warn_once("uverbs_write: process %d (%s) changed security contexts after opening file descriptor, this is not allowed.\n", 570 current->pid, current->comm); 571 return -EACCES; 572 } 573 574 if (count < sizeof(hdr)) 575 return -EINVAL; 576 577 if (copy_from_user(&hdr, buf, sizeof(hdr))) 578 return -EFAULT; 579 580 method_elm = uapi_get_method(uapi, hdr.command); 581 if (IS_ERR(method_elm)) 582 return PTR_ERR(method_elm); 583 584 if (method_elm->is_ex) { 585 if (count < (sizeof(hdr) + sizeof(ex_hdr))) 586 return -EINVAL; 587 if (copy_from_user(&ex_hdr, buf + sizeof(hdr), sizeof(ex_hdr))) 588 return -EFAULT; 589 } 590 591 ret = verify_hdr(&hdr, &ex_hdr, count, method_elm); 592 if (ret) 593 return ret; 594 595 srcu_key = srcu_read_lock(&file->device->disassociate_srcu); 596 597 buf += sizeof(hdr); 598 599 memset(bundle.attr_present, 0, sizeof(bundle.attr_present)); 600 bundle.ufile = file; 601 bundle.context = NULL; /* only valid if bundle has uobject */ 602 if (!method_elm->is_ex) { 603 size_t in_len = hdr.in_words * 4 - sizeof(hdr); 604 size_t out_len = hdr.out_words * 4; 605 u64 response = 0; 606 607 if (method_elm->has_udata) { 608 bundle.driver_udata.inlen = 609 in_len - method_elm->req_size; 610 in_len = method_elm->req_size; 611 if (bundle.driver_udata.inlen) 612 bundle.driver_udata.inbuf = buf + in_len; 613 else 614 bundle.driver_udata.inbuf = NULL; 615 } else { 616 memset(&bundle.driver_udata, 0, 617 sizeof(bundle.driver_udata)); 618 } 619 620 if (method_elm->has_resp) { 621 /* 622 * The macros check that if has_resp is set 623 * then the command request structure starts 624 * with a '__aligned u64 response' member. 625 */ 626 ret = get_user(response, (const u64 __user *)buf); 627 if (ret) 628 goto out_unlock; 629 630 if (method_elm->has_udata) { 631 bundle.driver_udata.outlen = 632 out_len - method_elm->resp_size; 633 out_len = method_elm->resp_size; 634 if (bundle.driver_udata.outlen) 635 bundle.driver_udata.outbuf = 636 u64_to_user_ptr(response + 637 out_len); 638 else 639 bundle.driver_udata.outbuf = NULL; 640 } 641 } else { 642 bundle.driver_udata.outlen = 0; 643 bundle.driver_udata.outbuf = NULL; 644 } 645 646 ib_uverbs_init_udata_buf_or_null( 647 &bundle.ucore, buf, u64_to_user_ptr(response), 648 in_len, out_len); 649 } else { 650 buf += sizeof(ex_hdr); 651 652 ib_uverbs_init_udata_buf_or_null(&bundle.ucore, buf, 653 u64_to_user_ptr(ex_hdr.response), 654 hdr.in_words * 8, hdr.out_words * 8); 655 656 ib_uverbs_init_udata_buf_or_null( 657 &bundle.driver_udata, buf + bundle.ucore.inlen, 658 u64_to_user_ptr(ex_hdr.response + bundle.ucore.outlen), 659 ex_hdr.provider_in_words * 8, 660 ex_hdr.provider_out_words * 8); 661 662 } 663 664 ret = method_elm->handler(&bundle); 665 out_unlock: 666 srcu_read_unlock(&file->device->disassociate_srcu, srcu_key); 667 return (ret) ? : count; 668 } 669 670 static const struct vm_operations_struct rdma_umap_ops; 671 672 static int ib_uverbs_mmap(struct file *filp, struct vm_area_struct *vma) 673 { 674 struct ib_uverbs_file *file = filp->private_data; 675 struct ib_ucontext *ucontext; 676 int ret = 0; 677 int srcu_key; 678 679 srcu_key = srcu_read_lock(&file->device->disassociate_srcu); 680 ucontext = ib_uverbs_get_ucontext_file(file); 681 if (IS_ERR(ucontext)) { 682 ret = PTR_ERR(ucontext); 683 goto out; 684 } 685 vma->vm_ops = &rdma_umap_ops; 686 ret = ucontext->device->mmap(ucontext, vma); 687 out: 688 srcu_read_unlock(&file->device->disassociate_srcu, srcu_key); 689 return ret; 690 } 691 692 /* 693 * The VMA has been dup'd, initialize the vm_private_data with a new tracking 694 * struct 695 */ 696 static void rdma_umap_open(struct vm_area_struct *vma) 697 { 698 struct ib_uverbs_file *ufile = vma->vm_file->private_data; 699 struct rdma_umap_priv *opriv = vma->vm_private_data; 700 struct rdma_umap_priv *priv; 701 702 if (!opriv) 703 return; 704 705 /* We are racing with disassociation */ 706 if (!down_read_trylock(&ufile->hw_destroy_rwsem)) 707 goto out_zap; 708 /* 709 * Disassociation already completed, the VMA should already be zapped. 710 */ 711 if (!ufile->ucontext) 712 goto out_unlock; 713 714 priv = kzalloc(sizeof(*priv), GFP_KERNEL); 715 if (!priv) 716 goto out_unlock; 717 rdma_umap_priv_init(priv, vma, opriv->entry); 718 719 up_read(&ufile->hw_destroy_rwsem); 720 return; 721 722 out_unlock: 723 up_read(&ufile->hw_destroy_rwsem); 724 out_zap: 725 /* 726 * We can't allow the VMA to be created with the actual IO pages, that 727 * would break our API contract, and it can't be stopped at this 728 * point, so zap it. 729 */ 730 vma->vm_private_data = NULL; 731 zap_vma_ptes(vma, vma->vm_start, vma->vm_end - vma->vm_start); 732 } 733 734 static void rdma_umap_close(struct vm_area_struct *vma) 735 { 736 struct ib_uverbs_file *ufile = vma->vm_file->private_data; 737 struct rdma_umap_priv *priv = vma->vm_private_data; 738 739 if (!priv) 740 return; 741 742 /* 743 * The vma holds a reference on the struct file that created it, which 744 * in turn means that the ib_uverbs_file is guaranteed to exist at 745 * this point. 746 */ 747 mutex_lock(&ufile->umap_lock); 748 749 list_del(&priv->list); 750 mutex_unlock(&ufile->umap_lock); 751 kfree(priv); 752 } 753 754 static const struct vm_operations_struct rdma_umap_ops = { 755 .open = rdma_umap_open, 756 .close = rdma_umap_close, 757 }; 758 759 void uverbs_user_mmap_disassociate(struct ib_uverbs_file *ufile) 760 { 761 struct rdma_umap_priv *priv, *next_priv; 762 763 lockdep_assert_held(&ufile->hw_destroy_rwsem); 764 765 while (1) { 766 struct mm_struct *mm = NULL; 767 768 /* Get an arbitrary mm pointer that hasn't been cleaned yet */ 769 mutex_lock(&ufile->umap_lock); 770 while (!list_empty(&ufile->umaps)) { 771 int ret; 772 773 priv = list_first_entry(&ufile->umaps, 774 struct rdma_umap_priv, list); 775 mm = priv->vma->vm_mm; 776 ret = mmget_not_zero(mm); 777 if (!ret) { 778 list_del_init(&priv->list); 779 if (priv->entry) { 780 rdma_user_mmap_entry_put(priv->entry); 781 priv->entry = NULL; 782 } 783 mm = NULL; 784 continue; 785 } 786 break; 787 } 788 mutex_unlock(&ufile->umap_lock); 789 if (!mm) 790 return; 791 792 /* 793 * The umap_lock is nested under mmap_sem since it used within 794 * the vma_ops callbacks, so we have to clean the list one mm 795 * at a time to get the lock ordering right. Typically there 796 * will only be one mm, so no big deal. 797 */ 798 down_read(&mm->mmap_sem); 799 mutex_lock(&ufile->umap_lock); 800 list_for_each_entry_safe (priv, next_priv, &ufile->umaps, 801 list) { 802 struct vm_area_struct *vma = priv->vma; 803 804 if (vma->vm_mm != mm) 805 continue; 806 list_del_init(&priv->list); 807 808 zap_vma_ptes(vma, vma->vm_start, 809 vma->vm_end - vma->vm_start); 810 } 811 mutex_unlock(&ufile->umap_lock); 812 up_read(&mm->mmap_sem); 813 mmput(mm); 814 } 815 } 816 817 /* 818 * ib_uverbs_open() does not need the BKL: 819 * 820 * - the ib_uverbs_device structures are properly reference counted and 821 * everything else is purely local to the file being created, so 822 * races against other open calls are not a problem; 823 * - there is no ioctl method to race against; 824 * - the open method will either immediately run -ENXIO, or all 825 * required initialization will be done. 826 */ 827 static int ib_uverbs_open(struct inode *inode, struct file *filp) 828 { 829 struct ib_uverbs_device *dev; 830 struct ib_uverbs_file *file; 831 struct ib_device *ib_dev; 832 int ret; 833 int module_dependent; 834 int srcu_key; 835 836 dev = container_of(inode->i_cdev->si_drv1, struct ib_uverbs_device, cdev); 837 if (!atomic_inc_not_zero(&dev->refcount)) 838 return -ENXIO; 839 840 get_device(&dev->dev); 841 srcu_key = srcu_read_lock(&dev->disassociate_srcu); 842 mutex_lock(&dev->lists_mutex); 843 ib_dev = srcu_dereference(dev->ib_dev, 844 &dev->disassociate_srcu); 845 if (!ib_dev) { 846 ret = -EIO; 847 goto err; 848 } 849 850 /* In case IB device supports disassociate ucontext, there is no hard 851 * dependency between uverbs device and its low level device. 852 */ 853 module_dependent = !(ib_dev->disassociate_ucontext); 854 855 if (module_dependent) { 856 if (!try_module_get(ib_dev->owner)) { 857 ret = -ENODEV; 858 goto err; 859 } 860 } 861 862 file = kzalloc(sizeof(*file), GFP_KERNEL); 863 if (!file) { 864 ret = -ENOMEM; 865 if (module_dependent) 866 goto err_module; 867 868 goto err; 869 } 870 871 file->device = dev; 872 kref_init(&file->ref); 873 mutex_init(&file->ucontext_lock); 874 875 spin_lock_init(&file->uobjects_lock); 876 INIT_LIST_HEAD(&file->uobjects); 877 init_rwsem(&file->hw_destroy_rwsem); 878 mutex_init(&file->umap_lock); 879 INIT_LIST_HEAD(&file->umaps); 880 881 filp->private_data = file; 882 list_add_tail(&file->list, &dev->uverbs_file_list); 883 mutex_unlock(&dev->lists_mutex); 884 srcu_read_unlock(&dev->disassociate_srcu, srcu_key); 885 886 setup_ufile_idr_uobject(file); 887 888 return nonseekable_open(inode, filp); 889 890 err_module: 891 module_put(ib_dev->owner); 892 893 err: 894 mutex_unlock(&dev->lists_mutex); 895 srcu_read_unlock(&dev->disassociate_srcu, srcu_key); 896 if (atomic_dec_and_test(&dev->refcount)) 897 ib_uverbs_comp_dev(dev); 898 899 put_device(&dev->dev); 900 return ret; 901 } 902 903 static int ib_uverbs_close(struct inode *inode, struct file *filp) 904 { 905 struct ib_uverbs_file *file = filp->private_data; 906 907 uverbs_destroy_ufile_hw(file, RDMA_REMOVE_CLOSE); 908 909 mutex_lock(&file->device->lists_mutex); 910 list_del_init(&file->list); 911 mutex_unlock(&file->device->lists_mutex); 912 913 kref_put(&file->ref, ib_uverbs_release_file); 914 915 return 0; 916 } 917 918 static const struct file_operations uverbs_fops = { 919 .owner = THIS_MODULE, 920 .write = ib_uverbs_write, 921 .open = ib_uverbs_open, 922 .release = ib_uverbs_close, 923 .llseek = no_llseek, 924 .unlocked_ioctl = ib_uverbs_ioctl, 925 .compat_ioctl = NULL, 926 }; 927 928 static const struct file_operations uverbs_mmap_fops = { 929 .owner = THIS_MODULE, 930 .write = ib_uverbs_write, 931 .mmap = ib_uverbs_mmap, 932 .open = ib_uverbs_open, 933 .release = ib_uverbs_close, 934 .llseek = no_llseek, 935 .unlocked_ioctl = ib_uverbs_ioctl, 936 .compat_ioctl = NULL, 937 }; 938 939 static struct ib_client uverbs_client = { 940 .name = "uverbs", 941 .add = ib_uverbs_add_one, 942 .remove = ib_uverbs_remove_one 943 }; 944 945 static ssize_t ibdev_show(struct device *device, struct device_attribute *attr, 946 char *buf) 947 { 948 struct ib_uverbs_device *dev = 949 container_of(device, struct ib_uverbs_device, dev); 950 int ret = -ENODEV; 951 int srcu_key; 952 struct ib_device *ib_dev; 953 954 srcu_key = srcu_read_lock(&dev->disassociate_srcu); 955 ib_dev = srcu_dereference(dev->ib_dev, &dev->disassociate_srcu); 956 if (ib_dev) 957 ret = sprintf(buf, "%s\n", ib_dev->name); 958 srcu_read_unlock(&dev->disassociate_srcu, srcu_key); 959 960 return ret; 961 } 962 static DEVICE_ATTR(ibdev, S_IRUGO, ibdev_show, NULL); 963 964 static ssize_t abi_version_show(struct device *device, 965 struct device_attribute *attr, char *buf) 966 { 967 struct ib_uverbs_device *dev = 968 container_of(device, struct ib_uverbs_device, dev); 969 int ret = -ENODEV; 970 int srcu_key; 971 struct ib_device *ib_dev; 972 973 srcu_key = srcu_read_lock(&dev->disassociate_srcu); 974 ib_dev = srcu_dereference(dev->ib_dev, &dev->disassociate_srcu); 975 if (ib_dev) 976 ret = sprintf(buf, "%d\n", ib_dev->uverbs_abi_ver); 977 srcu_read_unlock(&dev->disassociate_srcu, srcu_key); 978 979 return ret; 980 } 981 static DEVICE_ATTR(abi_version, S_IRUGO, abi_version_show, NULL); 982 983 static CLASS_ATTR_STRING(abi_version, S_IRUGO, 984 __stringify(IB_USER_VERBS_ABI_VERSION)); 985 986 static int ib_uverbs_create_uapi(struct ib_device *device, 987 struct ib_uverbs_device *uverbs_dev) 988 { 989 struct uverbs_api *uapi; 990 991 uapi = uverbs_alloc_api(device); 992 if (IS_ERR(uapi)) 993 return PTR_ERR(uapi); 994 995 uverbs_dev->uapi = uapi; 996 return 0; 997 } 998 999 static ssize_t 1000 show_dev_device(struct device *device, struct device_attribute *attr, char *buf) 1001 { 1002 struct ib_uverbs_device *dev = 1003 container_of(device, struct ib_uverbs_device, dev); 1004 1005 if (!dev->ib_dev->dma_device) 1006 return -ENODEV; 1007 1008 return sprintf(buf, "0x%04x\n", 1009 ((struct pci_dev *)dev->ib_dev->dma_device)->device); 1010 } 1011 static DEVICE_ATTR(device, S_IRUGO, show_dev_device, NULL); 1012 1013 static ssize_t 1014 show_dev_vendor(struct device *device, struct device_attribute *attr, char *buf) 1015 { 1016 struct ib_uverbs_device *dev = 1017 container_of(device, struct ib_uverbs_device, dev); 1018 1019 if (!dev->ib_dev->dma_device) 1020 return -ENODEV; 1021 1022 return sprintf(buf, "0x%04x\n", 1023 ((struct pci_dev *)dev->ib_dev->dma_device)->vendor); 1024 } 1025 static DEVICE_ATTR(vendor, S_IRUGO, show_dev_vendor, NULL); 1026 1027 static struct attribute *ib_dev_attrs[] = { 1028 &dev_attr_device.attr, 1029 &dev_attr_vendor.attr, 1030 NULL, 1031 }; 1032 1033 static const struct attribute_group dev_attr_group = { 1034 .name = "device", 1035 .attrs = ib_dev_attrs, 1036 }; 1037 1038 static void ib_uverbs_add_one(struct ib_device *device) 1039 { 1040 int devnum; 1041 dev_t base; 1042 struct ib_uverbs_device *uverbs_dev; 1043 int ret; 1044 1045 if (!device->alloc_ucontext) 1046 return; 1047 1048 uverbs_dev = kzalloc(sizeof(*uverbs_dev), GFP_KERNEL); 1049 if (!uverbs_dev) 1050 return; 1051 1052 ret = init_srcu_struct(&uverbs_dev->disassociate_srcu); 1053 if (ret) { 1054 kfree(uverbs_dev); 1055 return; 1056 } 1057 1058 uverbs_dev->dev.class = uverbs_class; 1059 uverbs_dev->dev.parent = device->dev.parent; 1060 uverbs_dev->dev.release = ib_uverbs_release_dev; 1061 device_initialize(&uverbs_dev->dev); 1062 atomic_set(&uverbs_dev->refcount, 1); 1063 init_completion(&uverbs_dev->comp); 1064 uverbs_dev->xrcd_tree = RB_ROOT; 1065 mutex_init(&uverbs_dev->xrcd_tree_mutex); 1066 mutex_init(&uverbs_dev->lists_mutex); 1067 INIT_LIST_HEAD(&uverbs_dev->uverbs_file_list); 1068 rcu_assign_pointer(uverbs_dev->ib_dev, device); 1069 uverbs_dev->num_comp_vectors = device->num_comp_vectors; 1070 1071 devnum = ida_alloc_max(&uverbs_ida, IB_UVERBS_MAX_DEVICES - 1, 1072 GFP_KERNEL); 1073 if (devnum < 0) 1074 goto err; 1075 uverbs_dev->devnum = devnum; 1076 if (devnum >= IB_UVERBS_NUM_FIXED_MINOR) 1077 base = dynamic_uverbs_dev + devnum - IB_UVERBS_NUM_FIXED_MINOR; 1078 else 1079 base = IB_UVERBS_BASE_DEV + devnum; 1080 1081 if (ib_uverbs_create_uapi(device, uverbs_dev)) 1082 goto err_uapi; 1083 1084 uverbs_dev->dev.devt = base; 1085 dev_set_name(&uverbs_dev->dev, "uverbs%d", uverbs_dev->devnum); 1086 1087 cdev_init(&uverbs_dev->cdev, 1088 device->mmap ? &uverbs_mmap_fops : &uverbs_fops); 1089 uverbs_dev->cdev.owner = THIS_MODULE; 1090 1091 kobject_set_name(&uverbs_dev->cdev.kobj, "uverbs%d", uverbs_dev->devnum); 1092 ret = cdev_device_add(&uverbs_dev->cdev, &uverbs_dev->dev); 1093 if (ret) 1094 goto err_uapi; 1095 1096 if (device_create_file(&uverbs_dev->dev, &dev_attr_ibdev)) 1097 goto err_cdev; 1098 if (device_create_file(&uverbs_dev->dev, &dev_attr_abi_version)) 1099 goto err_cdev; 1100 if (sysfs_create_group(&uverbs_dev->dev.kobj, &dev_attr_group)) 1101 goto err_cdev; 1102 1103 ib_set_client_data(device, &uverbs_client, uverbs_dev); 1104 return; 1105 1106 err_cdev: 1107 cdev_device_del(&uverbs_dev->cdev, &uverbs_dev->dev); 1108 err_uapi: 1109 ida_free(&uverbs_ida, devnum); 1110 err: 1111 if (atomic_dec_and_test(&uverbs_dev->refcount)) 1112 ib_uverbs_comp_dev(uverbs_dev); 1113 wait_for_completion(&uverbs_dev->comp); 1114 put_device(&uverbs_dev->dev); 1115 return; 1116 } 1117 1118 static void ib_uverbs_free_hw_resources(struct ib_uverbs_device *uverbs_dev, 1119 struct ib_device *ib_dev) 1120 { 1121 struct ib_uverbs_file *file; 1122 1123 /* Pending running commands to terminate */ 1124 uverbs_disassociate_api_pre(uverbs_dev); 1125 1126 mutex_lock(&uverbs_dev->lists_mutex); 1127 while (!list_empty(&uverbs_dev->uverbs_file_list)) { 1128 file = list_first_entry(&uverbs_dev->uverbs_file_list, 1129 struct ib_uverbs_file, list); 1130 list_del_init(&file->list); 1131 kref_get(&file->ref); 1132 1133 /* We must release the mutex before going ahead and calling 1134 * uverbs_cleanup_ufile, as it might end up indirectly calling 1135 * uverbs_close, for example due to freeing the resources (e.g 1136 * mmput). 1137 */ 1138 mutex_unlock(&uverbs_dev->lists_mutex); 1139 1140 ib_uverbs_async_handler(READ_ONCE(file->async_file), 0, 1141 IB_EVENT_DEVICE_FATAL, NULL, NULL); 1142 1143 uverbs_destroy_ufile_hw(file, RDMA_REMOVE_DRIVER_REMOVE); 1144 kref_put(&file->ref, ib_uverbs_release_file); 1145 1146 mutex_lock(&uverbs_dev->lists_mutex); 1147 } 1148 mutex_unlock(&uverbs_dev->lists_mutex); 1149 1150 uverbs_disassociate_api(uverbs_dev->uapi); 1151 } 1152 1153 static void ib_uverbs_remove_one(struct ib_device *device, void *client_data) 1154 { 1155 struct ib_uverbs_device *uverbs_dev = client_data; 1156 int wait_clients = 1; 1157 1158 if (!uverbs_dev) 1159 return; 1160 1161 cdev_device_del(&uverbs_dev->cdev, &uverbs_dev->dev); 1162 ida_free(&uverbs_ida, uverbs_dev->devnum); 1163 1164 if (device->disassociate_ucontext) { 1165 /* We disassociate HW resources and immediately return. 1166 * Userspace will see a EIO errno for all future access. 1167 * Upon returning, ib_device may be freed internally and is not 1168 * valid any more. 1169 * uverbs_device is still available until all clients close 1170 * their files, then the uverbs device ref count will be zero 1171 * and its resources will be freed. 1172 * Note: At this point no more files can be opened since the 1173 * cdev was deleted, however active clients can still issue 1174 * commands and close their open files. 1175 */ 1176 ib_uverbs_free_hw_resources(uverbs_dev, device); 1177 wait_clients = 0; 1178 } 1179 1180 if (atomic_dec_and_test(&uverbs_dev->refcount)) 1181 ib_uverbs_comp_dev(uverbs_dev); 1182 if (wait_clients) 1183 wait_for_completion(&uverbs_dev->comp); 1184 1185 put_device(&uverbs_dev->dev); 1186 } 1187 1188 static char *uverbs_devnode(struct device *dev, umode_t *mode) 1189 { 1190 if (mode) 1191 *mode = 0666; 1192 return kasprintf(GFP_KERNEL, "infiniband/%s", dev_name(dev)); 1193 } 1194 1195 static int __init ib_uverbs_init(void) 1196 { 1197 int ret; 1198 1199 ret = register_chrdev_region(IB_UVERBS_BASE_DEV, 1200 IB_UVERBS_NUM_FIXED_MINOR, 1201 "infiniband_verbs"); 1202 if (ret) { 1203 pr_err("user_verbs: couldn't register device number\n"); 1204 goto out; 1205 } 1206 1207 ret = alloc_chrdev_region(&dynamic_uverbs_dev, 0, 1208 IB_UVERBS_NUM_DYNAMIC_MINOR, 1209 "infiniband_verbs"); 1210 if (ret) { 1211 pr_err("couldn't register dynamic device number\n"); 1212 goto out_alloc; 1213 } 1214 1215 uverbs_class = class_create(THIS_MODULE, "infiniband_verbs"); 1216 if (IS_ERR(uverbs_class)) { 1217 ret = PTR_ERR(uverbs_class); 1218 pr_err("user_verbs: couldn't create class infiniband_verbs\n"); 1219 goto out_chrdev; 1220 } 1221 1222 uverbs_class->devnode = uverbs_devnode; 1223 1224 ret = class_create_file(uverbs_class, &class_attr_abi_version.attr); 1225 if (ret) { 1226 pr_err("user_verbs: couldn't create abi_version attribute\n"); 1227 goto out_class; 1228 } 1229 1230 ret = ib_register_client(&uverbs_client); 1231 if (ret) { 1232 pr_err("user_verbs: couldn't register client\n"); 1233 goto out_class; 1234 } 1235 1236 return 0; 1237 1238 out_class: 1239 class_destroy(uverbs_class); 1240 1241 out_chrdev: 1242 unregister_chrdev_region(dynamic_uverbs_dev, 1243 IB_UVERBS_NUM_DYNAMIC_MINOR); 1244 1245 out_alloc: 1246 unregister_chrdev_region(IB_UVERBS_BASE_DEV, 1247 IB_UVERBS_NUM_FIXED_MINOR); 1248 1249 out: 1250 return ret; 1251 } 1252 1253 static void __exit ib_uverbs_cleanup(void) 1254 { 1255 ib_unregister_client(&uverbs_client); 1256 class_destroy(uverbs_class); 1257 unregister_chrdev_region(IB_UVERBS_BASE_DEV, 1258 IB_UVERBS_NUM_FIXED_MINOR); 1259 unregister_chrdev_region(dynamic_uverbs_dev, 1260 IB_UVERBS_NUM_DYNAMIC_MINOR); 1261 } 1262 1263 module_init_order(ib_uverbs_init, SI_ORDER_FIFTH); 1264 module_exit_order(ib_uverbs_cleanup, SI_ORDER_FIFTH); 1265