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 60MODULE_AUTHOR("Roland Dreier"); 61MODULE_DESCRIPTION("InfiniBand userspace verbs access"); 62MODULE_LICENSE("Dual BSD/GPL"); 63 64enum { 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 74static dev_t dynamic_uverbs_dev; 75 76static DEFINE_IDA(uverbs_ida); 77static int ib_uverbs_add_one(struct ib_device *device); 78static void ib_uverbs_remove_one(struct ib_device *device, void *client_data); 79 80static char *uverbs_devnode(const struct device *dev, umode_t *mode) 81{ 82 if (mode) 83 *mode = 0666; 84 return kasprintf(GFP_KERNEL, "infiniband/%s", dev_name(dev)); 85} 86 87static const struct class uverbs_class = { 88 .name = "infiniband_verbs", 89 .devnode = uverbs_devnode, 90}; 91 92/* 93 * Must be called with the ufile->device->disassociate_srcu held, and the lock 94 * must be held until use of the ucontext is finished. 95 */ 96struct ib_ucontext *ib_uverbs_get_ucontext_file(struct ib_uverbs_file *ufile) 97{ 98 /* 99 * We do not hold the hw_destroy_rwsem lock for this flow, instead 100 * srcu is used. It does not matter if someone races this with 101 * get_context, we get NULL or valid ucontext. 102 */ 103 struct ib_ucontext *ucontext = smp_load_acquire(&ufile->ucontext); 104 105 if (!srcu_dereference(ufile->device->ib_dev, 106 &ufile->device->disassociate_srcu)) 107 return ERR_PTR(-EIO); 108 109 if (!ucontext) 110 return ERR_PTR(-EINVAL); 111 112 return ucontext; 113} 114EXPORT_SYMBOL(ib_uverbs_get_ucontext_file); 115 116int uverbs_dealloc_mw(struct ib_mw *mw) 117{ 118 struct ib_pd *pd = mw->pd; 119 int ret; 120 121 ret = mw->device->ops.dealloc_mw(mw); 122 if (ret) 123 return ret; 124 125 atomic_dec(&pd->usecnt); 126 kfree(mw); 127 return ret; 128} 129 130static void ib_uverbs_release_dev(struct device *device) 131{ 132 struct ib_uverbs_device *dev = 133 container_of(device, struct ib_uverbs_device, dev); 134 135 uverbs_destroy_api(dev->uapi); 136 cleanup_srcu_struct(&dev->disassociate_srcu); 137 mutex_destroy(&dev->lists_mutex); 138 mutex_destroy(&dev->xrcd_tree_mutex); 139 kfree(dev); 140} 141 142void ib_uverbs_release_ucq(struct ib_uverbs_completion_event_file *ev_file, 143 struct ib_ucq_object *uobj) 144{ 145 struct ib_uverbs_event *evt, *tmp; 146 147 if (ev_file) { 148 spin_lock_irq(&ev_file->ev_queue.lock); 149 list_for_each_entry_safe(evt, tmp, &uobj->comp_list, obj_list) { 150 list_del(&evt->list); 151 kfree(evt); 152 } 153 spin_unlock_irq(&ev_file->ev_queue.lock); 154 155 uverbs_uobject_put(&ev_file->uobj); 156 } 157 158 ib_uverbs_release_uevent(&uobj->uevent); 159} 160 161void ib_uverbs_release_uevent(struct ib_uevent_object *uobj) 162{ 163 struct ib_uverbs_async_event_file *async_file = uobj->event_file; 164 struct ib_uverbs_event *evt, *tmp; 165 166 if (!async_file) 167 return; 168 169 spin_lock_irq(&async_file->ev_queue.lock); 170 list_for_each_entry_safe(evt, tmp, &uobj->event_list, obj_list) { 171 list_del(&evt->list); 172 kfree(evt); 173 } 174 spin_unlock_irq(&async_file->ev_queue.lock); 175 uverbs_uobject_put(&async_file->uobj); 176} 177 178void ib_uverbs_detach_umcast(struct ib_qp *qp, 179 struct ib_uqp_object *uobj) 180{ 181 struct ib_uverbs_mcast_entry *mcast, *tmp; 182 183 list_for_each_entry_safe(mcast, tmp, &uobj->mcast_list, list) { 184 ib_detach_mcast(qp, &mcast->gid, mcast->lid); 185 list_del(&mcast->list); 186 kfree(mcast); 187 } 188} 189 190static void ib_uverbs_comp_dev(struct ib_uverbs_device *dev) 191{ 192 complete(&dev->comp); 193} 194 195void ib_uverbs_release_file(struct kref *ref) 196{ 197 struct ib_uverbs_file *file = 198 container_of(ref, struct ib_uverbs_file, ref); 199 struct ib_device *ib_dev; 200 int srcu_key; 201 202 release_ufile_idr_uobject(file); 203 204 srcu_key = srcu_read_lock(&file->device->disassociate_srcu); 205 ib_dev = srcu_dereference(file->device->ib_dev, 206 &file->device->disassociate_srcu); 207 if (ib_dev && !ib_dev->ops.disassociate_ucontext) 208 module_put(ib_dev->ops.owner); 209 srcu_read_unlock(&file->device->disassociate_srcu, srcu_key); 210 211 if (refcount_dec_and_test(&file->device->refcount)) 212 ib_uverbs_comp_dev(file->device); 213 214 if (file->default_async_file) 215 uverbs_uobject_put(&file->default_async_file->uobj); 216 put_device(&file->device->dev); 217 218 if (file->disassociate_page) 219 __free_pages(file->disassociate_page, 0); 220 mutex_destroy(&file->umap_lock); 221 mutex_destroy(&file->ucontext_lock); 222 kfree(file); 223} 224 225static ssize_t ib_uverbs_event_read(struct ib_uverbs_event_queue *ev_queue, 226 struct file *filp, char __user *buf, 227 size_t count, loff_t *pos, 228 size_t eventsz) 229{ 230 struct ib_uverbs_event *event; 231 int ret = 0; 232 233 spin_lock_irq(&ev_queue->lock); 234 235 while (list_empty(&ev_queue->event_list)) { 236 if (ev_queue->is_closed) { 237 spin_unlock_irq(&ev_queue->lock); 238 return -EIO; 239 } 240 241 spin_unlock_irq(&ev_queue->lock); 242 if (filp->f_flags & O_NONBLOCK) 243 return -EAGAIN; 244 245 if (wait_event_interruptible(ev_queue->poll_wait, 246 (!list_empty(&ev_queue->event_list) || 247 ev_queue->is_closed))) 248 return -ERESTARTSYS; 249 250 spin_lock_irq(&ev_queue->lock); 251 } 252 253 event = list_entry(ev_queue->event_list.next, struct ib_uverbs_event, list); 254 255 if (eventsz > count) { 256 ret = -EINVAL; 257 event = NULL; 258 } else { 259 list_del(ev_queue->event_list.next); 260 if (event->counter) { 261 ++(*event->counter); 262 list_del(&event->obj_list); 263 } 264 } 265 266 spin_unlock_irq(&ev_queue->lock); 267 268 if (event) { 269 if (copy_to_user(buf, event, eventsz)) 270 ret = -EFAULT; 271 else 272 ret = eventsz; 273 } 274 275 kfree(event); 276 277 return ret; 278} 279 280static ssize_t ib_uverbs_async_event_read(struct file *filp, char __user *buf, 281 size_t count, loff_t *pos) 282{ 283 struct ib_uverbs_async_event_file *file = filp->private_data; 284 285 return ib_uverbs_event_read(&file->ev_queue, filp, buf, count, pos, 286 sizeof(struct ib_uverbs_async_event_desc)); 287} 288 289static ssize_t ib_uverbs_comp_event_read(struct file *filp, char __user *buf, 290 size_t count, loff_t *pos) 291{ 292 struct ib_uverbs_completion_event_file *comp_ev_file = 293 filp->private_data; 294 295 return ib_uverbs_event_read(&comp_ev_file->ev_queue, filp, buf, count, 296 pos, 297 sizeof(struct ib_uverbs_comp_event_desc)); 298} 299 300static __poll_t ib_uverbs_event_poll(struct ib_uverbs_event_queue *ev_queue, 301 struct file *filp, 302 struct poll_table_struct *wait) 303{ 304 __poll_t pollflags = 0; 305 306 poll_wait(filp, &ev_queue->poll_wait, wait); 307 308 spin_lock_irq(&ev_queue->lock); 309 if (!list_empty(&ev_queue->event_list)) 310 pollflags = EPOLLIN | EPOLLRDNORM; 311 else if (ev_queue->is_closed) 312 pollflags = EPOLLERR; 313 spin_unlock_irq(&ev_queue->lock); 314 315 return pollflags; 316} 317 318static __poll_t ib_uverbs_async_event_poll(struct file *filp, 319 struct poll_table_struct *wait) 320{ 321 struct ib_uverbs_async_event_file *file = filp->private_data; 322 323 return ib_uverbs_event_poll(&file->ev_queue, filp, wait); 324} 325 326static __poll_t ib_uverbs_comp_event_poll(struct file *filp, 327 struct poll_table_struct *wait) 328{ 329 struct ib_uverbs_completion_event_file *comp_ev_file = 330 filp->private_data; 331 332 return ib_uverbs_event_poll(&comp_ev_file->ev_queue, filp, wait); 333} 334 335static int ib_uverbs_async_event_fasync(int fd, struct file *filp, int on) 336{ 337 struct ib_uverbs_async_event_file *file = filp->private_data; 338 339 return fasync_helper(fd, filp, on, &file->ev_queue.async_queue); 340} 341 342static int ib_uverbs_comp_event_fasync(int fd, struct file *filp, int on) 343{ 344 struct ib_uverbs_completion_event_file *comp_ev_file = 345 filp->private_data; 346 347 return fasync_helper(fd, filp, on, &comp_ev_file->ev_queue.async_queue); 348} 349 350const struct file_operations uverbs_event_fops = { 351 .owner = THIS_MODULE, 352 .read = ib_uverbs_comp_event_read, 353 .poll = ib_uverbs_comp_event_poll, 354 .release = uverbs_uobject_fd_release, 355 .fasync = ib_uverbs_comp_event_fasync, 356 .llseek = no_llseek, 357}; 358 359const struct file_operations uverbs_async_event_fops = { 360 .owner = THIS_MODULE, 361 .read = ib_uverbs_async_event_read, 362 .poll = ib_uverbs_async_event_poll, 363 .release = uverbs_async_event_release, 364 .fasync = ib_uverbs_async_event_fasync, 365 .llseek = no_llseek, 366}; 367 368void ib_uverbs_comp_handler(struct ib_cq *cq, void *cq_context) 369{ 370 struct ib_uverbs_event_queue *ev_queue = cq_context; 371 struct ib_ucq_object *uobj; 372 struct ib_uverbs_event *entry; 373 unsigned long flags; 374 375 if (!ev_queue) 376 return; 377 378 spin_lock_irqsave(&ev_queue->lock, flags); 379 if (ev_queue->is_closed) { 380 spin_unlock_irqrestore(&ev_queue->lock, flags); 381 return; 382 } 383 384 entry = kmalloc(sizeof(*entry), GFP_ATOMIC); 385 if (!entry) { 386 spin_unlock_irqrestore(&ev_queue->lock, flags); 387 return; 388 } 389 390 uobj = cq->uobject; 391 392 entry->desc.comp.cq_handle = cq->uobject->uevent.uobject.user_handle; 393 entry->counter = &uobj->comp_events_reported; 394 395 list_add_tail(&entry->list, &ev_queue->event_list); 396 list_add_tail(&entry->obj_list, &uobj->comp_list); 397 spin_unlock_irqrestore(&ev_queue->lock, flags); 398 399 wake_up_interruptible(&ev_queue->poll_wait); 400 kill_fasync(&ev_queue->async_queue, SIGIO, POLL_IN); 401} 402 403void ib_uverbs_async_handler(struct ib_uverbs_async_event_file *async_file, 404 __u64 element, __u64 event, 405 struct list_head *obj_list, u32 *counter) 406{ 407 struct ib_uverbs_event *entry; 408 unsigned long flags; 409 410 if (!async_file) 411 return; 412 413 spin_lock_irqsave(&async_file->ev_queue.lock, flags); 414 if (async_file->ev_queue.is_closed) { 415 spin_unlock_irqrestore(&async_file->ev_queue.lock, flags); 416 return; 417 } 418 419 entry = kmalloc(sizeof(*entry), GFP_ATOMIC); 420 if (!entry) { 421 spin_unlock_irqrestore(&async_file->ev_queue.lock, flags); 422 return; 423 } 424 425 entry->desc.async.element = element; 426 entry->desc.async.event_type = event; 427 entry->desc.async.reserved = 0; 428 entry->counter = counter; 429 430 list_add_tail(&entry->list, &async_file->ev_queue.event_list); 431 if (obj_list) 432 list_add_tail(&entry->obj_list, obj_list); 433 spin_unlock_irqrestore(&async_file->ev_queue.lock, flags); 434 435 wake_up_interruptible(&async_file->ev_queue.poll_wait); 436 kill_fasync(&async_file->ev_queue.async_queue, SIGIO, POLL_IN); 437} 438 439static void uverbs_uobj_event(struct ib_uevent_object *eobj, 440 struct ib_event *event) 441{ 442 ib_uverbs_async_handler(eobj->event_file, 443 eobj->uobject.user_handle, event->event, 444 &eobj->event_list, &eobj->events_reported); 445} 446 447void ib_uverbs_cq_event_handler(struct ib_event *event, void *context_ptr) 448{ 449 uverbs_uobj_event(&event->element.cq->uobject->uevent, event); 450} 451 452void ib_uverbs_qp_event_handler(struct ib_event *event, void *context_ptr) 453{ 454 /* for XRC target qp's, check that qp is live */ 455 if (!event->element.qp->uobject) 456 return; 457 458 uverbs_uobj_event(&event->element.qp->uobject->uevent, event); 459} 460 461void ib_uverbs_wq_event_handler(struct ib_event *event, void *context_ptr) 462{ 463 uverbs_uobj_event(&event->element.wq->uobject->uevent, event); 464} 465 466void ib_uverbs_srq_event_handler(struct ib_event *event, void *context_ptr) 467{ 468 uverbs_uobj_event(&event->element.srq->uobject->uevent, event); 469} 470 471static void ib_uverbs_event_handler(struct ib_event_handler *handler, 472 struct ib_event *event) 473{ 474 ib_uverbs_async_handler( 475 container_of(handler, struct ib_uverbs_async_event_file, 476 event_handler), 477 event->element.port_num, event->event, NULL, NULL); 478} 479 480void ib_uverbs_init_event_queue(struct ib_uverbs_event_queue *ev_queue) 481{ 482 spin_lock_init(&ev_queue->lock); 483 INIT_LIST_HEAD(&ev_queue->event_list); 484 init_waitqueue_head(&ev_queue->poll_wait); 485 ev_queue->is_closed = 0; 486 ev_queue->async_queue = NULL; 487} 488 489void ib_uverbs_init_async_event_file( 490 struct ib_uverbs_async_event_file *async_file) 491{ 492 struct ib_uverbs_file *uverbs_file = async_file->uobj.ufile; 493 struct ib_device *ib_dev = async_file->uobj.context->device; 494 495 ib_uverbs_init_event_queue(&async_file->ev_queue); 496 497 /* The first async_event_file becomes the default one for the file. */ 498 mutex_lock(&uverbs_file->ucontext_lock); 499 if (!uverbs_file->default_async_file) { 500 /* Pairs with the put in ib_uverbs_release_file */ 501 uverbs_uobject_get(&async_file->uobj); 502 smp_store_release(&uverbs_file->default_async_file, async_file); 503 } 504 mutex_unlock(&uverbs_file->ucontext_lock); 505 506 INIT_IB_EVENT_HANDLER(&async_file->event_handler, ib_dev, 507 ib_uverbs_event_handler); 508 ib_register_event_handler(&async_file->event_handler); 509} 510 511static ssize_t verify_hdr(struct ib_uverbs_cmd_hdr *hdr, 512 struct ib_uverbs_ex_cmd_hdr *ex_hdr, size_t count, 513 const struct uverbs_api_write_method *method_elm) 514{ 515 if (method_elm->is_ex) { 516 count -= sizeof(*hdr) + sizeof(*ex_hdr); 517 518 if ((hdr->in_words + ex_hdr->provider_in_words) * 8 != count) 519 return -EINVAL; 520 521 if (hdr->in_words * 8 < method_elm->req_size) 522 return -ENOSPC; 523 524 if (ex_hdr->cmd_hdr_reserved) 525 return -EINVAL; 526 527 if (ex_hdr->response) { 528 if (!hdr->out_words && !ex_hdr->provider_out_words) 529 return -EINVAL; 530 531 if (hdr->out_words * 8 < method_elm->resp_size) 532 return -ENOSPC; 533 534 if (!access_ok(u64_to_user_ptr(ex_hdr->response), 535 (hdr->out_words + ex_hdr->provider_out_words) * 8)) 536 return -EFAULT; 537 } else { 538 if (hdr->out_words || ex_hdr->provider_out_words) 539 return -EINVAL; 540 } 541 542 return 0; 543 } 544 545 /* not extended command */ 546 if (hdr->in_words * 4 != count) 547 return -EINVAL; 548 549 if (count < method_elm->req_size + sizeof(*hdr)) { 550 /* 551 * rdma-core v18 and v19 have a bug where they send DESTROY_CQ 552 * with a 16 byte write instead of 24. Old kernels didn't 553 * check the size so they allowed this. Now that the size is 554 * checked provide a compatibility work around to not break 555 * those userspaces. 556 */ 557 if (hdr->command == IB_USER_VERBS_CMD_DESTROY_CQ && 558 count == 16) { 559 hdr->in_words = 6; 560 return 0; 561 } 562 return -ENOSPC; 563 } 564 if (hdr->out_words * 4 < method_elm->resp_size) 565 return -ENOSPC; 566 567 return 0; 568} 569 570static ssize_t ib_uverbs_write(struct file *filp, const char __user *buf, 571 size_t count, loff_t *pos) 572{ 573 struct ib_uverbs_file *file = filp->private_data; 574 const struct uverbs_api_write_method *method_elm; 575 struct uverbs_api *uapi = file->device->uapi; 576 struct ib_uverbs_ex_cmd_hdr ex_hdr; 577 struct ib_uverbs_cmd_hdr hdr; 578 struct uverbs_attr_bundle bundle; 579 int srcu_key; 580 ssize_t ret; 581 582 if (!ib_safe_file_access(filp)) { 583 pr_err_once("uverbs_write: process %d (%s) changed security contexts after opening file descriptor, this is not allowed.\n", 584 task_tgid_vnr(current), current->comm); 585 return -EACCES; 586 } 587 588 if (count < sizeof(hdr)) 589 return -EINVAL; 590 591 if (copy_from_user(&hdr, buf, sizeof(hdr))) 592 return -EFAULT; 593 594 method_elm = uapi_get_method(uapi, hdr.command); 595 if (IS_ERR(method_elm)) 596 return PTR_ERR(method_elm); 597 598 if (method_elm->is_ex) { 599 if (count < (sizeof(hdr) + sizeof(ex_hdr))) 600 return -EINVAL; 601 if (copy_from_user(&ex_hdr, buf + sizeof(hdr), sizeof(ex_hdr))) 602 return -EFAULT; 603 } 604 605 ret = verify_hdr(&hdr, &ex_hdr, count, method_elm); 606 if (ret) 607 return ret; 608 609 srcu_key = srcu_read_lock(&file->device->disassociate_srcu); 610 611 buf += sizeof(hdr); 612 613 memset(bundle.attr_present, 0, sizeof(bundle.attr_present)); 614 bundle.ufile = file; 615 bundle.context = NULL; /* only valid if bundle has uobject */ 616 bundle.uobject = NULL; 617 if (!method_elm->is_ex) { 618 size_t in_len = hdr.in_words * 4 - sizeof(hdr); 619 size_t out_len = hdr.out_words * 4; 620 u64 response = 0; 621 622 if (method_elm->has_udata) { 623 bundle.driver_udata.inlen = 624 in_len - method_elm->req_size; 625 in_len = method_elm->req_size; 626 if (bundle.driver_udata.inlen) 627 bundle.driver_udata.inbuf = buf + in_len; 628 else 629 bundle.driver_udata.inbuf = NULL; 630 } else { 631 memset(&bundle.driver_udata, 0, 632 sizeof(bundle.driver_udata)); 633 } 634 635 if (method_elm->has_resp) { 636 /* 637 * The macros check that if has_resp is set 638 * then the command request structure starts 639 * with a '__aligned u64 response' member. 640 */ 641 ret = get_user(response, (const u64 __user *)buf); 642 if (ret) 643 goto out_unlock; 644 645 if (method_elm->has_udata) { 646 bundle.driver_udata.outlen = 647 out_len - method_elm->resp_size; 648 out_len = method_elm->resp_size; 649 if (bundle.driver_udata.outlen) 650 bundle.driver_udata.outbuf = 651 u64_to_user_ptr(response + 652 out_len); 653 else 654 bundle.driver_udata.outbuf = NULL; 655 } 656 } else { 657 bundle.driver_udata.outlen = 0; 658 bundle.driver_udata.outbuf = NULL; 659 } 660 661 ib_uverbs_init_udata_buf_or_null( 662 &bundle.ucore, buf, u64_to_user_ptr(response), 663 in_len, out_len); 664 } else { 665 buf += sizeof(ex_hdr); 666 667 ib_uverbs_init_udata_buf_or_null(&bundle.ucore, buf, 668 u64_to_user_ptr(ex_hdr.response), 669 hdr.in_words * 8, hdr.out_words * 8); 670 671 ib_uverbs_init_udata_buf_or_null( 672 &bundle.driver_udata, buf + bundle.ucore.inlen, 673 u64_to_user_ptr(ex_hdr.response) + bundle.ucore.outlen, 674 ex_hdr.provider_in_words * 8, 675 ex_hdr.provider_out_words * 8); 676 677 } 678 679 ret = method_elm->handler(&bundle); 680 if (bundle.uobject) 681 uverbs_finalize_object(bundle.uobject, UVERBS_ACCESS_NEW, true, 682 !ret, &bundle); 683out_unlock: 684 srcu_read_unlock(&file->device->disassociate_srcu, srcu_key); 685 return (ret) ? : count; 686} 687 688static const struct vm_operations_struct rdma_umap_ops; 689 690static int ib_uverbs_mmap(struct file *filp, struct vm_area_struct *vma) 691{ 692 struct ib_uverbs_file *file = filp->private_data; 693 struct ib_ucontext *ucontext; 694 int ret = 0; 695 int srcu_key; 696 697 srcu_key = srcu_read_lock(&file->device->disassociate_srcu); 698 ucontext = ib_uverbs_get_ucontext_file(file); 699 if (IS_ERR(ucontext)) { 700 ret = PTR_ERR(ucontext); 701 goto out; 702 } 703 vma->vm_ops = &rdma_umap_ops; 704 ret = ucontext->device->ops.mmap(ucontext, vma); 705out: 706 srcu_read_unlock(&file->device->disassociate_srcu, srcu_key); 707 return ret; 708} 709 710/* 711 * The VMA has been dup'd, initialize the vm_private_data with a new tracking 712 * struct 713 */ 714static void rdma_umap_open(struct vm_area_struct *vma) 715{ 716 struct ib_uverbs_file *ufile = vma->vm_file->private_data; 717 struct rdma_umap_priv *opriv = vma->vm_private_data; 718 struct rdma_umap_priv *priv; 719 720 if (!opriv) 721 return; 722 723 /* We are racing with disassociation */ 724 if (!down_read_trylock(&ufile->hw_destroy_rwsem)) 725 goto out_zap; 726 /* 727 * Disassociation already completed, the VMA should already be zapped. 728 */ 729 if (!ufile->ucontext) 730 goto out_unlock; 731 732 priv = kzalloc(sizeof(*priv), GFP_KERNEL); 733 if (!priv) 734 goto out_unlock; 735 rdma_umap_priv_init(priv, vma, opriv->entry); 736 737 up_read(&ufile->hw_destroy_rwsem); 738 return; 739 740out_unlock: 741 up_read(&ufile->hw_destroy_rwsem); 742out_zap: 743 /* 744 * We can't allow the VMA to be created with the actual IO pages, that 745 * would break our API contract, and it can't be stopped at this 746 * point, so zap it. 747 */ 748 vma->vm_private_data = NULL; 749 zap_vma_ptes(vma, vma->vm_start, vma->vm_end - vma->vm_start); 750} 751 752static void rdma_umap_close(struct vm_area_struct *vma) 753{ 754 struct ib_uverbs_file *ufile = vma->vm_file->private_data; 755 struct rdma_umap_priv *priv = vma->vm_private_data; 756 757 if (!priv) 758 return; 759 760 /* 761 * The vma holds a reference on the struct file that created it, which 762 * in turn means that the ib_uverbs_file is guaranteed to exist at 763 * this point. 764 */ 765 mutex_lock(&ufile->umap_lock); 766 if (priv->entry) 767 rdma_user_mmap_entry_put(priv->entry); 768 769 list_del(&priv->list); 770 mutex_unlock(&ufile->umap_lock); 771 kfree(priv); 772} 773 774/* 775 * Once the zap_vma_ptes has been called touches to the VMA will come here and 776 * we return a dummy writable zero page for all the pfns. 777 */ 778static vm_fault_t rdma_umap_fault(struct vm_fault *vmf) 779{ 780 struct ib_uverbs_file *ufile = vmf->vma->vm_file->private_data; 781 struct rdma_umap_priv *priv = vmf->vma->vm_private_data; 782 vm_fault_t ret = 0; 783 784 if (!priv) 785 return VM_FAULT_SIGBUS; 786 787 /* Read only pages can just use the system zero page. */ 788 if (!(vmf->vma->vm_flags & (VM_WRITE | VM_MAYWRITE))) { 789 vmf->page = ZERO_PAGE(vmf->address); 790 get_page(vmf->page); 791 return 0; 792 } 793 794 mutex_lock(&ufile->umap_lock); 795 if (!ufile->disassociate_page) 796 ufile->disassociate_page = 797 alloc_pages(vmf->gfp_mask | __GFP_ZERO, 0); 798 799 if (ufile->disassociate_page) { 800 /* 801 * This VMA is forced to always be shared so this doesn't have 802 * to worry about COW. 803 */ 804 vmf->page = ufile->disassociate_page; 805 get_page(vmf->page); 806 } else { 807 ret = VM_FAULT_SIGBUS; 808 } 809 mutex_unlock(&ufile->umap_lock); 810 811 return ret; 812} 813 814static const struct vm_operations_struct rdma_umap_ops = { 815 .open = rdma_umap_open, 816 .close = rdma_umap_close, 817 .fault = rdma_umap_fault, 818}; 819 820void uverbs_user_mmap_disassociate(struct ib_uverbs_file *ufile) 821{ 822 struct rdma_umap_priv *priv, *next_priv; 823 824 lockdep_assert_held(&ufile->hw_destroy_rwsem); 825 826 while (1) { 827 struct mm_struct *mm = NULL; 828 829 /* Get an arbitrary mm pointer that hasn't been cleaned yet */ 830 mutex_lock(&ufile->umap_lock); 831 while (!list_empty(&ufile->umaps)) { 832 int ret; 833 834 priv = list_first_entry(&ufile->umaps, 835 struct rdma_umap_priv, list); 836 mm = priv->vma->vm_mm; 837 ret = mmget_not_zero(mm); 838 if (!ret) { 839 list_del_init(&priv->list); 840 if (priv->entry) { 841 rdma_user_mmap_entry_put(priv->entry); 842 priv->entry = NULL; 843 } 844 mm = NULL; 845 continue; 846 } 847 break; 848 } 849 mutex_unlock(&ufile->umap_lock); 850 if (!mm) 851 return; 852 853 /* 854 * The umap_lock is nested under mmap_lock since it used within 855 * the vma_ops callbacks, so we have to clean the list one mm 856 * at a time to get the lock ordering right. Typically there 857 * will only be one mm, so no big deal. 858 */ 859 mmap_read_lock(mm); 860 mutex_lock(&ufile->umap_lock); 861 list_for_each_entry_safe (priv, next_priv, &ufile->umaps, 862 list) { 863 struct vm_area_struct *vma = priv->vma; 864 865 if (vma->vm_mm != mm) 866 continue; 867 list_del_init(&priv->list); 868 869 zap_vma_ptes(vma, vma->vm_start, 870 vma->vm_end - vma->vm_start); 871 872 if (priv->entry) { 873 rdma_user_mmap_entry_put(priv->entry); 874 priv->entry = NULL; 875 } 876 } 877 mutex_unlock(&ufile->umap_lock); 878 mmap_read_unlock(mm); 879 mmput(mm); 880 } 881} 882 883/* 884 * ib_uverbs_open() does not need the BKL: 885 * 886 * - the ib_uverbs_device structures are properly reference counted and 887 * everything else is purely local to the file being created, so 888 * races against other open calls are not a problem; 889 * - there is no ioctl method to race against; 890 * - the open method will either immediately run -ENXIO, or all 891 * required initialization will be done. 892 */ 893static int ib_uverbs_open(struct inode *inode, struct file *filp) 894{ 895 struct ib_uverbs_device *dev; 896 struct ib_uverbs_file *file; 897 struct ib_device *ib_dev; 898 int ret; 899 int module_dependent; 900 int srcu_key; 901 902 dev = container_of(inode->i_cdev, struct ib_uverbs_device, cdev); 903 if (!refcount_inc_not_zero(&dev->refcount)) 904 return -ENXIO; 905 906 get_device(&dev->dev); 907 srcu_key = srcu_read_lock(&dev->disassociate_srcu); 908 mutex_lock(&dev->lists_mutex); 909 ib_dev = srcu_dereference(dev->ib_dev, 910 &dev->disassociate_srcu); 911 if (!ib_dev) { 912 ret = -EIO; 913 goto err; 914 } 915 916 if (!rdma_dev_access_netns(ib_dev, current->nsproxy->net_ns)) { 917 ret = -EPERM; 918 goto err; 919 } 920 921 /* In case IB device supports disassociate ucontext, there is no hard 922 * dependency between uverbs device and its low level device. 923 */ 924 module_dependent = !(ib_dev->ops.disassociate_ucontext); 925 926 if (module_dependent) { 927 if (!try_module_get(ib_dev->ops.owner)) { 928 ret = -ENODEV; 929 goto err; 930 } 931 } 932 933 file = kzalloc(sizeof(*file), GFP_KERNEL); 934 if (!file) { 935 ret = -ENOMEM; 936 if (module_dependent) 937 goto err_module; 938 939 goto err; 940 } 941 942 file->device = dev; 943 kref_init(&file->ref); 944 mutex_init(&file->ucontext_lock); 945 946 spin_lock_init(&file->uobjects_lock); 947 INIT_LIST_HEAD(&file->uobjects); 948 init_rwsem(&file->hw_destroy_rwsem); 949 mutex_init(&file->umap_lock); 950 INIT_LIST_HEAD(&file->umaps); 951 952 filp->private_data = file; 953 list_add_tail(&file->list, &dev->uverbs_file_list); 954 mutex_unlock(&dev->lists_mutex); 955 srcu_read_unlock(&dev->disassociate_srcu, srcu_key); 956 957 setup_ufile_idr_uobject(file); 958 959 return stream_open(inode, filp); 960 961err_module: 962 module_put(ib_dev->ops.owner); 963 964err: 965 mutex_unlock(&dev->lists_mutex); 966 srcu_read_unlock(&dev->disassociate_srcu, srcu_key); 967 if (refcount_dec_and_test(&dev->refcount)) 968 ib_uverbs_comp_dev(dev); 969 970 put_device(&dev->dev); 971 return ret; 972} 973 974static int ib_uverbs_close(struct inode *inode, struct file *filp) 975{ 976 struct ib_uverbs_file *file = filp->private_data; 977 978 uverbs_destroy_ufile_hw(file, RDMA_REMOVE_CLOSE); 979 980 mutex_lock(&file->device->lists_mutex); 981 list_del_init(&file->list); 982 mutex_unlock(&file->device->lists_mutex); 983 984 kref_put(&file->ref, ib_uverbs_release_file); 985 986 return 0; 987} 988 989static const struct file_operations uverbs_fops = { 990 .owner = THIS_MODULE, 991 .write = ib_uverbs_write, 992 .open = ib_uverbs_open, 993 .release = ib_uverbs_close, 994 .llseek = no_llseek, 995 .unlocked_ioctl = ib_uverbs_ioctl, 996 .compat_ioctl = compat_ptr_ioctl, 997}; 998 999static const struct file_operations uverbs_mmap_fops = { 1000 .owner = THIS_MODULE, 1001 .write = ib_uverbs_write, 1002 .mmap = ib_uverbs_mmap, 1003 .open = ib_uverbs_open, 1004 .release = ib_uverbs_close, 1005 .llseek = no_llseek, 1006 .unlocked_ioctl = ib_uverbs_ioctl, 1007 .compat_ioctl = compat_ptr_ioctl, 1008}; 1009 1010static int ib_uverbs_get_nl_info(struct ib_device *ibdev, void *client_data, 1011 struct ib_client_nl_info *res) 1012{ 1013 struct ib_uverbs_device *uverbs_dev = client_data; 1014 int ret; 1015 1016 if (res->port != -1) 1017 return -EINVAL; 1018 1019 res->abi = ibdev->ops.uverbs_abi_ver; 1020 res->cdev = &uverbs_dev->dev; 1021 1022 /* 1023 * To support DRIVER_ID binding in userspace some of the driver need 1024 * upgrading to expose their PCI dependent revision information 1025 * through get_context instead of relying on modalias matching. When 1026 * the drivers are fixed they can drop this flag. 1027 */ 1028 if (!ibdev->ops.uverbs_no_driver_id_binding) { 1029 ret = nla_put_u32(res->nl_msg, RDMA_NLDEV_ATTR_UVERBS_DRIVER_ID, 1030 ibdev->ops.driver_id); 1031 if (ret) 1032 return ret; 1033 } 1034 return 0; 1035} 1036 1037static struct ib_client uverbs_client = { 1038 .name = "uverbs", 1039 .no_kverbs_req = true, 1040 .add = ib_uverbs_add_one, 1041 .remove = ib_uverbs_remove_one, 1042 .get_nl_info = ib_uverbs_get_nl_info, 1043}; 1044MODULE_ALIAS_RDMA_CLIENT("uverbs"); 1045 1046static ssize_t ibdev_show(struct device *device, struct device_attribute *attr, 1047 char *buf) 1048{ 1049 struct ib_uverbs_device *dev = 1050 container_of(device, struct ib_uverbs_device, dev); 1051 int ret = -ENODEV; 1052 int srcu_key; 1053 struct ib_device *ib_dev; 1054 1055 srcu_key = srcu_read_lock(&dev->disassociate_srcu); 1056 ib_dev = srcu_dereference(dev->ib_dev, &dev->disassociate_srcu); 1057 if (ib_dev) 1058 ret = sysfs_emit(buf, "%s\n", dev_name(&ib_dev->dev)); 1059 srcu_read_unlock(&dev->disassociate_srcu, srcu_key); 1060 1061 return ret; 1062} 1063static DEVICE_ATTR_RO(ibdev); 1064 1065static ssize_t abi_version_show(struct device *device, 1066 struct device_attribute *attr, char *buf) 1067{ 1068 struct ib_uverbs_device *dev = 1069 container_of(device, struct ib_uverbs_device, dev); 1070 int ret = -ENODEV; 1071 int srcu_key; 1072 struct ib_device *ib_dev; 1073 1074 srcu_key = srcu_read_lock(&dev->disassociate_srcu); 1075 ib_dev = srcu_dereference(dev->ib_dev, &dev->disassociate_srcu); 1076 if (ib_dev) 1077 ret = sysfs_emit(buf, "%u\n", ib_dev->ops.uverbs_abi_ver); 1078 srcu_read_unlock(&dev->disassociate_srcu, srcu_key); 1079 1080 return ret; 1081} 1082static DEVICE_ATTR_RO(abi_version); 1083 1084static struct attribute *ib_dev_attrs[] = { 1085 &dev_attr_abi_version.attr, 1086 &dev_attr_ibdev.attr, 1087 NULL, 1088}; 1089 1090static const struct attribute_group dev_attr_group = { 1091 .attrs = ib_dev_attrs, 1092}; 1093 1094static CLASS_ATTR_STRING(abi_version, S_IRUGO, 1095 __stringify(IB_USER_VERBS_ABI_VERSION)); 1096 1097static int ib_uverbs_create_uapi(struct ib_device *device, 1098 struct ib_uverbs_device *uverbs_dev) 1099{ 1100 struct uverbs_api *uapi; 1101 1102 uapi = uverbs_alloc_api(device); 1103 if (IS_ERR(uapi)) 1104 return PTR_ERR(uapi); 1105 1106 uverbs_dev->uapi = uapi; 1107 return 0; 1108} 1109 1110static int ib_uverbs_add_one(struct ib_device *device) 1111{ 1112 int devnum; 1113 dev_t base; 1114 struct ib_uverbs_device *uverbs_dev; 1115 int ret; 1116 1117 if (!device->ops.alloc_ucontext) 1118 return -EOPNOTSUPP; 1119 1120 uverbs_dev = kzalloc(sizeof(*uverbs_dev), GFP_KERNEL); 1121 if (!uverbs_dev) 1122 return -ENOMEM; 1123 1124 ret = init_srcu_struct(&uverbs_dev->disassociate_srcu); 1125 if (ret) { 1126 kfree(uverbs_dev); 1127 return -ENOMEM; 1128 } 1129 1130 device_initialize(&uverbs_dev->dev); 1131 uverbs_dev->dev.class = &uverbs_class; 1132 uverbs_dev->dev.parent = device->dev.parent; 1133 uverbs_dev->dev.release = ib_uverbs_release_dev; 1134 uverbs_dev->groups[0] = &dev_attr_group; 1135 uverbs_dev->dev.groups = uverbs_dev->groups; 1136 refcount_set(&uverbs_dev->refcount, 1); 1137 init_completion(&uverbs_dev->comp); 1138 uverbs_dev->xrcd_tree = RB_ROOT; 1139 mutex_init(&uverbs_dev->xrcd_tree_mutex); 1140 mutex_init(&uverbs_dev->lists_mutex); 1141 INIT_LIST_HEAD(&uverbs_dev->uverbs_file_list); 1142 rcu_assign_pointer(uverbs_dev->ib_dev, device); 1143 uverbs_dev->num_comp_vectors = device->num_comp_vectors; 1144 1145 devnum = ida_alloc_max(&uverbs_ida, IB_UVERBS_MAX_DEVICES - 1, 1146 GFP_KERNEL); 1147 if (devnum < 0) { 1148 ret = -ENOMEM; 1149 goto err; 1150 } 1151 uverbs_dev->devnum = devnum; 1152 if (devnum >= IB_UVERBS_NUM_FIXED_MINOR) 1153 base = dynamic_uverbs_dev + devnum - IB_UVERBS_NUM_FIXED_MINOR; 1154 else 1155 base = IB_UVERBS_BASE_DEV + devnum; 1156 1157 ret = ib_uverbs_create_uapi(device, uverbs_dev); 1158 if (ret) 1159 goto err_uapi; 1160 1161 uverbs_dev->dev.devt = base; 1162 dev_set_name(&uverbs_dev->dev, "uverbs%d", uverbs_dev->devnum); 1163 1164 cdev_init(&uverbs_dev->cdev, 1165 device->ops.mmap ? &uverbs_mmap_fops : &uverbs_fops); 1166 uverbs_dev->cdev.owner = THIS_MODULE; 1167 1168 ret = cdev_device_add(&uverbs_dev->cdev, &uverbs_dev->dev); 1169 if (ret) 1170 goto err_uapi; 1171 1172 ib_set_client_data(device, &uverbs_client, uverbs_dev); 1173 return 0; 1174 1175err_uapi: 1176 ida_free(&uverbs_ida, devnum); 1177err: 1178 if (refcount_dec_and_test(&uverbs_dev->refcount)) 1179 ib_uverbs_comp_dev(uverbs_dev); 1180 wait_for_completion(&uverbs_dev->comp); 1181 put_device(&uverbs_dev->dev); 1182 return ret; 1183} 1184 1185static void ib_uverbs_free_hw_resources(struct ib_uverbs_device *uverbs_dev, 1186 struct ib_device *ib_dev) 1187{ 1188 struct ib_uverbs_file *file; 1189 1190 /* Pending running commands to terminate */ 1191 uverbs_disassociate_api_pre(uverbs_dev); 1192 1193 mutex_lock(&uverbs_dev->lists_mutex); 1194 while (!list_empty(&uverbs_dev->uverbs_file_list)) { 1195 file = list_first_entry(&uverbs_dev->uverbs_file_list, 1196 struct ib_uverbs_file, list); 1197 list_del_init(&file->list); 1198 kref_get(&file->ref); 1199 1200 /* We must release the mutex before going ahead and calling 1201 * uverbs_cleanup_ufile, as it might end up indirectly calling 1202 * uverbs_close, for example due to freeing the resources (e.g 1203 * mmput). 1204 */ 1205 mutex_unlock(&uverbs_dev->lists_mutex); 1206 1207 uverbs_destroy_ufile_hw(file, RDMA_REMOVE_DRIVER_REMOVE); 1208 kref_put(&file->ref, ib_uverbs_release_file); 1209 1210 mutex_lock(&uverbs_dev->lists_mutex); 1211 } 1212 mutex_unlock(&uverbs_dev->lists_mutex); 1213 1214 uverbs_disassociate_api(uverbs_dev->uapi); 1215} 1216 1217static void ib_uverbs_remove_one(struct ib_device *device, void *client_data) 1218{ 1219 struct ib_uverbs_device *uverbs_dev = client_data; 1220 int wait_clients = 1; 1221 1222 cdev_device_del(&uverbs_dev->cdev, &uverbs_dev->dev); 1223 ida_free(&uverbs_ida, uverbs_dev->devnum); 1224 1225 if (device->ops.disassociate_ucontext) { 1226 /* We disassociate HW resources and immediately return. 1227 * Userspace will see a EIO errno for all future access. 1228 * Upon returning, ib_device may be freed internally and is not 1229 * valid any more. 1230 * uverbs_device is still available until all clients close 1231 * their files, then the uverbs device ref count will be zero 1232 * and its resources will be freed. 1233 * Note: At this point no more files can be opened since the 1234 * cdev was deleted, however active clients can still issue 1235 * commands and close their open files. 1236 */ 1237 ib_uverbs_free_hw_resources(uverbs_dev, device); 1238 wait_clients = 0; 1239 } 1240 1241 if (refcount_dec_and_test(&uverbs_dev->refcount)) 1242 ib_uverbs_comp_dev(uverbs_dev); 1243 if (wait_clients) 1244 wait_for_completion(&uverbs_dev->comp); 1245 1246 put_device(&uverbs_dev->dev); 1247} 1248 1249static int __init ib_uverbs_init(void) 1250{ 1251 int ret; 1252 1253 ret = register_chrdev_region(IB_UVERBS_BASE_DEV, 1254 IB_UVERBS_NUM_FIXED_MINOR, 1255 "infiniband_verbs"); 1256 if (ret) { 1257 pr_err("user_verbs: couldn't register device number\n"); 1258 goto out; 1259 } 1260 1261 ret = alloc_chrdev_region(&dynamic_uverbs_dev, 0, 1262 IB_UVERBS_NUM_DYNAMIC_MINOR, 1263 "infiniband_verbs"); 1264 if (ret) { 1265 pr_err("couldn't register dynamic device number\n"); 1266 goto out_alloc; 1267 } 1268 1269 ret = class_register(&uverbs_class); 1270 if (ret) { 1271 pr_err("user_verbs: couldn't create class infiniband_verbs\n"); 1272 goto out_chrdev; 1273 } 1274 1275 ret = class_create_file(&uverbs_class, &class_attr_abi_version.attr); 1276 if (ret) { 1277 pr_err("user_verbs: couldn't create abi_version attribute\n"); 1278 goto out_class; 1279 } 1280 1281 ret = ib_register_client(&uverbs_client); 1282 if (ret) { 1283 pr_err("user_verbs: couldn't register client\n"); 1284 goto out_class; 1285 } 1286 1287 return 0; 1288 1289out_class: 1290 class_unregister(&uverbs_class); 1291 1292out_chrdev: 1293 unregister_chrdev_region(dynamic_uverbs_dev, 1294 IB_UVERBS_NUM_DYNAMIC_MINOR); 1295 1296out_alloc: 1297 unregister_chrdev_region(IB_UVERBS_BASE_DEV, 1298 IB_UVERBS_NUM_FIXED_MINOR); 1299 1300out: 1301 return ret; 1302} 1303 1304static void __exit ib_uverbs_cleanup(void) 1305{ 1306 ib_unregister_client(&uverbs_client); 1307 class_unregister(&uverbs_class); 1308 unregister_chrdev_region(IB_UVERBS_BASE_DEV, 1309 IB_UVERBS_NUM_FIXED_MINOR); 1310 unregister_chrdev_region(dynamic_uverbs_dev, 1311 IB_UVERBS_NUM_DYNAMIC_MINOR); 1312 mmu_notifier_synchronize(); 1313} 1314 1315module_init(ib_uverbs_init); 1316module_exit(ib_uverbs_cleanup); 1317