1 /* 2 * Copyright (c) 2005 Topspin Communications. All rights reserved. 3 * Copyright (c) 2005, 2006 Cisco Systems. All rights reserved. 4 * Copyright (c) 2005 Mellanox Technologies. All rights reserved. 5 * Copyright (c) 2005 Voltaire, Inc. All rights reserved. 6 * Copyright (c) 2005 PathScale, Inc. All rights reserved. 7 * 8 * This software is available to you under a choice of one of two 9 * licenses. You may choose to be licensed under the terms of the GNU 10 * General Public License (GPL) Version 2, available from the file 11 * COPYING in the main directory of this source tree, or the 12 * OpenIB.org BSD license below: 13 * 14 * Redistribution and use in source and binary forms, with or 15 * without modification, are permitted provided that the following 16 * conditions are met: 17 * 18 * - Redistributions of source code must retain the above 19 * copyright notice, this list of conditions and the following 20 * disclaimer. 21 * 22 * - Redistributions in binary form must reproduce the above 23 * copyright notice, this list of conditions and the following 24 * disclaimer in the documentation and/or other materials 25 * provided with the distribution. 26 * 27 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, 28 * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF 29 * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND 30 * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS 31 * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN 32 * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN 33 * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE 34 * SOFTWARE. 35 */ 36 37 #include <linux/module.h> 38 #include <linux/init.h> 39 #include <linux/device.h> 40 #include <linux/err.h> 41 #include <linux/fs.h> 42 #include <linux/poll.h> 43 #include <linux/sched.h> 44 #include <linux/file.h> 45 #include <linux/cdev.h> 46 #include <linux/anon_inodes.h> 47 #include <linux/slab.h> 48 #include <linux/sched/mm.h> 49 50 #include <linux/uaccess.h> 51 52 #include <rdma/ib.h> 53 #include <rdma/uverbs_std_types.h> 54 #include <rdma/rdma_netlink.h> 55 56 #include "uverbs.h" 57 #include "core_priv.h" 58 #include "rdma_core.h" 59 60 MODULE_AUTHOR("Roland Dreier"); 61 MODULE_DESCRIPTION("InfiniBand userspace verbs access"); 62 MODULE_LICENSE("Dual BSD/GPL"); 63 64 enum { 65 IB_UVERBS_MAJOR = 231, 66 IB_UVERBS_BASE_MINOR = 192, 67 IB_UVERBS_MAX_DEVICES = RDMA_MAX_PORTS, 68 IB_UVERBS_NUM_FIXED_MINOR = 32, 69 IB_UVERBS_NUM_DYNAMIC_MINOR = IB_UVERBS_MAX_DEVICES - IB_UVERBS_NUM_FIXED_MINOR, 70 }; 71 72 #define IB_UVERBS_BASE_DEV MKDEV(IB_UVERBS_MAJOR, IB_UVERBS_BASE_MINOR) 73 74 static dev_t dynamic_uverbs_dev; 75 static struct class *uverbs_class; 76 77 static DEFINE_IDA(uverbs_ida); 78 static int ib_uverbs_add_one(struct ib_device *device); 79 static void ib_uverbs_remove_one(struct ib_device *device, void *client_data); 80 81 /* 82 * Must be called with the ufile->device->disassociate_srcu held, and the lock 83 * must be held until use of the ucontext is finished. 84 */ 85 struct ib_ucontext *ib_uverbs_get_ucontext_file(struct ib_uverbs_file *ufile) 86 { 87 /* 88 * We do not hold the hw_destroy_rwsem lock for this flow, instead 89 * srcu is used. It does not matter if someone races this with 90 * get_context, we get NULL or valid ucontext. 91 */ 92 struct ib_ucontext *ucontext = smp_load_acquire(&ufile->ucontext); 93 94 if (!srcu_dereference(ufile->device->ib_dev, 95 &ufile->device->disassociate_srcu)) 96 return ERR_PTR(-EIO); 97 98 if (!ucontext) 99 return ERR_PTR(-EINVAL); 100 101 return ucontext; 102 } 103 EXPORT_SYMBOL(ib_uverbs_get_ucontext_file); 104 105 int uverbs_dealloc_mw(struct ib_mw *mw) 106 { 107 struct ib_pd *pd = mw->pd; 108 int ret; 109 110 ret = mw->device->ops.dealloc_mw(mw); 111 if (ret) 112 return ret; 113 114 atomic_dec(&pd->usecnt); 115 kfree(mw); 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 = uobj->event_file; 153 struct ib_uverbs_event *evt, *tmp; 154 155 if (!async_file) 156 return; 157 158 spin_lock_irq(&async_file->ev_queue.lock); 159 list_for_each_entry_safe(evt, tmp, &uobj->event_list, obj_list) { 160 list_del(&evt->list); 161 kfree(evt); 162 } 163 spin_unlock_irq(&async_file->ev_queue.lock); 164 uverbs_uobject_put(&async_file->uobj); 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->ops.disassociate_ucontext) 197 module_put(ib_dev->ops.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->default_async_file) 204 uverbs_uobject_put(&file->default_async_file->uobj); 205 put_device(&file->device->dev); 206 207 if (file->disassociate_page) 208 __free_pages(file->disassociate_page, 0); 209 mutex_destroy(&file->umap_lock); 210 mutex_destroy(&file->ucontext_lock); 211 kfree(file); 212 } 213 214 static ssize_t ib_uverbs_event_read(struct ib_uverbs_event_queue *ev_queue, 215 struct file *filp, char __user *buf, 216 size_t count, loff_t *pos, 217 size_t eventsz) 218 { 219 struct ib_uverbs_event *event; 220 int ret = 0; 221 222 spin_lock_irq(&ev_queue->lock); 223 224 while (list_empty(&ev_queue->event_list)) { 225 spin_unlock_irq(&ev_queue->lock); 226 227 if (filp->f_flags & O_NONBLOCK) 228 return -EAGAIN; 229 230 if (wait_event_interruptible(ev_queue->poll_wait, 231 (!list_empty(&ev_queue->event_list) || 232 ev_queue->is_closed))) 233 return -ERESTARTSYS; 234 235 spin_lock_irq(&ev_queue->lock); 236 237 /* If device was disassociated and no event exists set an error */ 238 if (list_empty(&ev_queue->event_list) && ev_queue->is_closed) { 239 spin_unlock_irq(&ev_queue->lock); 240 return -EIO; 241 } 242 } 243 244 event = list_entry(ev_queue->event_list.next, struct ib_uverbs_event, list); 245 246 if (eventsz > count) { 247 ret = -EINVAL; 248 event = NULL; 249 } else { 250 list_del(ev_queue->event_list.next); 251 if (event->counter) { 252 ++(*event->counter); 253 list_del(&event->obj_list); 254 } 255 } 256 257 spin_unlock_irq(&ev_queue->lock); 258 259 if (event) { 260 if (copy_to_user(buf, event, eventsz)) 261 ret = -EFAULT; 262 else 263 ret = eventsz; 264 } 265 266 kfree(event); 267 268 return ret; 269 } 270 271 static ssize_t ib_uverbs_async_event_read(struct file *filp, char __user *buf, 272 size_t count, loff_t *pos) 273 { 274 struct ib_uverbs_async_event_file *file = filp->private_data; 275 276 return ib_uverbs_event_read(&file->ev_queue, filp, buf, count, pos, 277 sizeof(struct ib_uverbs_async_event_desc)); 278 } 279 280 static ssize_t ib_uverbs_comp_event_read(struct file *filp, char __user *buf, 281 size_t count, loff_t *pos) 282 { 283 struct ib_uverbs_completion_event_file *comp_ev_file = 284 filp->private_data; 285 286 return ib_uverbs_event_read(&comp_ev_file->ev_queue, filp, buf, count, 287 pos, 288 sizeof(struct ib_uverbs_comp_event_desc)); 289 } 290 291 static __poll_t ib_uverbs_event_poll(struct ib_uverbs_event_queue *ev_queue, 292 struct file *filp, 293 struct poll_table_struct *wait) 294 { 295 __poll_t pollflags = 0; 296 297 poll_wait(filp, &ev_queue->poll_wait, wait); 298 299 spin_lock_irq(&ev_queue->lock); 300 if (!list_empty(&ev_queue->event_list)) 301 pollflags = EPOLLIN | EPOLLRDNORM; 302 else if (ev_queue->is_closed) 303 pollflags = EPOLLERR; 304 spin_unlock_irq(&ev_queue->lock); 305 306 return pollflags; 307 } 308 309 static __poll_t ib_uverbs_async_event_poll(struct file *filp, 310 struct poll_table_struct *wait) 311 { 312 struct ib_uverbs_async_event_file *file = filp->private_data; 313 314 return ib_uverbs_event_poll(&file->ev_queue, filp, wait); 315 } 316 317 static __poll_t ib_uverbs_comp_event_poll(struct file *filp, 318 struct poll_table_struct *wait) 319 { 320 struct ib_uverbs_completion_event_file *comp_ev_file = 321 filp->private_data; 322 323 return ib_uverbs_event_poll(&comp_ev_file->ev_queue, filp, wait); 324 } 325 326 static int ib_uverbs_async_event_fasync(int fd, struct file *filp, int on) 327 { 328 struct ib_uverbs_async_event_file *file = filp->private_data; 329 330 return fasync_helper(fd, filp, on, &file->ev_queue.async_queue); 331 } 332 333 static int ib_uverbs_comp_event_fasync(int fd, struct file *filp, int on) 334 { 335 struct ib_uverbs_completion_event_file *comp_ev_file = 336 filp->private_data; 337 338 return fasync_helper(fd, filp, on, &comp_ev_file->ev_queue.async_queue); 339 } 340 341 const struct file_operations uverbs_event_fops = { 342 .owner = THIS_MODULE, 343 .read = ib_uverbs_comp_event_read, 344 .poll = ib_uverbs_comp_event_poll, 345 .release = uverbs_uobject_fd_release, 346 .fasync = ib_uverbs_comp_event_fasync, 347 .llseek = no_llseek, 348 }; 349 350 const struct file_operations uverbs_async_event_fops = { 351 .owner = THIS_MODULE, 352 .read = ib_uverbs_async_event_read, 353 .poll = ib_uverbs_async_event_poll, 354 .release = uverbs_async_event_release, 355 .fasync = ib_uverbs_async_event_fasync, 356 .llseek = no_llseek, 357 }; 358 359 void ib_uverbs_comp_handler(struct ib_cq *cq, void *cq_context) 360 { 361 struct ib_uverbs_event_queue *ev_queue = cq_context; 362 struct ib_ucq_object *uobj; 363 struct ib_uverbs_event *entry; 364 unsigned long flags; 365 366 if (!ev_queue) 367 return; 368 369 spin_lock_irqsave(&ev_queue->lock, flags); 370 if (ev_queue->is_closed) { 371 spin_unlock_irqrestore(&ev_queue->lock, flags); 372 return; 373 } 374 375 entry = kmalloc(sizeof(*entry), GFP_ATOMIC); 376 if (!entry) { 377 spin_unlock_irqrestore(&ev_queue->lock, flags); 378 return; 379 } 380 381 uobj = cq->uobject; 382 383 entry->desc.comp.cq_handle = cq->uobject->uevent.uobject.user_handle; 384 entry->counter = &uobj->comp_events_reported; 385 386 list_add_tail(&entry->list, &ev_queue->event_list); 387 list_add_tail(&entry->obj_list, &uobj->comp_list); 388 spin_unlock_irqrestore(&ev_queue->lock, flags); 389 390 wake_up_interruptible(&ev_queue->poll_wait); 391 kill_fasync(&ev_queue->async_queue, SIGIO, POLL_IN); 392 } 393 394 void ib_uverbs_async_handler(struct ib_uverbs_async_event_file *async_file, 395 __u64 element, __u64 event, 396 struct list_head *obj_list, u32 *counter) 397 { 398 struct ib_uverbs_event *entry; 399 unsigned long flags; 400 401 if (!async_file) 402 return; 403 404 spin_lock_irqsave(&async_file->ev_queue.lock, flags); 405 if (async_file->ev_queue.is_closed) { 406 spin_unlock_irqrestore(&async_file->ev_queue.lock, flags); 407 return; 408 } 409 410 entry = kmalloc(sizeof(*entry), GFP_ATOMIC); 411 if (!entry) { 412 spin_unlock_irqrestore(&async_file->ev_queue.lock, flags); 413 return; 414 } 415 416 entry->desc.async.element = element; 417 entry->desc.async.event_type = event; 418 entry->desc.async.reserved = 0; 419 entry->counter = counter; 420 421 list_add_tail(&entry->list, &async_file->ev_queue.event_list); 422 if (obj_list) 423 list_add_tail(&entry->obj_list, obj_list); 424 spin_unlock_irqrestore(&async_file->ev_queue.lock, flags); 425 426 wake_up_interruptible(&async_file->ev_queue.poll_wait); 427 kill_fasync(&async_file->ev_queue.async_queue, SIGIO, POLL_IN); 428 } 429 430 static void uverbs_uobj_event(struct ib_uevent_object *eobj, 431 struct ib_event *event) 432 { 433 ib_uverbs_async_handler(eobj->event_file, 434 eobj->uobject.user_handle, event->event, 435 &eobj->event_list, &eobj->events_reported); 436 } 437 438 void ib_uverbs_cq_event_handler(struct ib_event *event, void *context_ptr) 439 { 440 uverbs_uobj_event(&event->element.cq->uobject->uevent, event); 441 } 442 443 void ib_uverbs_qp_event_handler(struct ib_event *event, void *context_ptr) 444 { 445 /* for XRC target qp's, check that qp is live */ 446 if (!event->element.qp->uobject) 447 return; 448 449 uverbs_uobj_event(&event->element.qp->uobject->uevent, event); 450 } 451 452 void ib_uverbs_wq_event_handler(struct ib_event *event, void *context_ptr) 453 { 454 uverbs_uobj_event(&event->element.wq->uobject->uevent, event); 455 } 456 457 void ib_uverbs_srq_event_handler(struct ib_event *event, void *context_ptr) 458 { 459 uverbs_uobj_event(&event->element.srq->uobject->uevent, event); 460 } 461 462 static void ib_uverbs_event_handler(struct ib_event_handler *handler, 463 struct ib_event *event) 464 { 465 ib_uverbs_async_handler( 466 container_of(handler, struct ib_uverbs_async_event_file, 467 event_handler), 468 event->element.port_num, event->event, NULL, NULL); 469 } 470 471 void ib_uverbs_init_event_queue(struct ib_uverbs_event_queue *ev_queue) 472 { 473 spin_lock_init(&ev_queue->lock); 474 INIT_LIST_HEAD(&ev_queue->event_list); 475 init_waitqueue_head(&ev_queue->poll_wait); 476 ev_queue->is_closed = 0; 477 ev_queue->async_queue = NULL; 478 } 479 480 void ib_uverbs_init_async_event_file( 481 struct ib_uverbs_async_event_file *async_file) 482 { 483 struct ib_uverbs_file *uverbs_file = async_file->uobj.ufile; 484 struct ib_device *ib_dev = async_file->uobj.context->device; 485 486 ib_uverbs_init_event_queue(&async_file->ev_queue); 487 488 /* The first async_event_file becomes the default one for the file. */ 489 mutex_lock(&uverbs_file->ucontext_lock); 490 if (!uverbs_file->default_async_file) { 491 /* Pairs with the put in ib_uverbs_release_file */ 492 uverbs_uobject_get(&async_file->uobj); 493 smp_store_release(&uverbs_file->default_async_file, async_file); 494 } 495 mutex_unlock(&uverbs_file->ucontext_lock); 496 497 INIT_IB_EVENT_HANDLER(&async_file->event_handler, ib_dev, 498 ib_uverbs_event_handler); 499 ib_register_event_handler(&async_file->event_handler); 500 } 501 502 static ssize_t verify_hdr(struct ib_uverbs_cmd_hdr *hdr, 503 struct ib_uverbs_ex_cmd_hdr *ex_hdr, size_t count, 504 const struct uverbs_api_write_method *method_elm) 505 { 506 if (method_elm->is_ex) { 507 count -= sizeof(*hdr) + sizeof(*ex_hdr); 508 509 if ((hdr->in_words + ex_hdr->provider_in_words) * 8 != count) 510 return -EINVAL; 511 512 if (hdr->in_words * 8 < method_elm->req_size) 513 return -ENOSPC; 514 515 if (ex_hdr->cmd_hdr_reserved) 516 return -EINVAL; 517 518 if (ex_hdr->response) { 519 if (!hdr->out_words && !ex_hdr->provider_out_words) 520 return -EINVAL; 521 522 if (hdr->out_words * 8 < method_elm->resp_size) 523 return -ENOSPC; 524 525 if (!access_ok(u64_to_user_ptr(ex_hdr->response), 526 (hdr->out_words + ex_hdr->provider_out_words) * 8)) 527 return -EFAULT; 528 } else { 529 if (hdr->out_words || ex_hdr->provider_out_words) 530 return -EINVAL; 531 } 532 533 return 0; 534 } 535 536 /* not extended command */ 537 if (hdr->in_words * 4 != count) 538 return -EINVAL; 539 540 if (count < method_elm->req_size + sizeof(hdr)) { 541 /* 542 * rdma-core v18 and v19 have a bug where they send DESTROY_CQ 543 * with a 16 byte write instead of 24. Old kernels didn't 544 * check the size so they allowed this. Now that the size is 545 * checked provide a compatibility work around to not break 546 * those userspaces. 547 */ 548 if (hdr->command == IB_USER_VERBS_CMD_DESTROY_CQ && 549 count == 16) { 550 hdr->in_words = 6; 551 return 0; 552 } 553 return -ENOSPC; 554 } 555 if (hdr->out_words * 4 < method_elm->resp_size) 556 return -ENOSPC; 557 558 return 0; 559 } 560 561 static ssize_t ib_uverbs_write(struct file *filp, const char __user *buf, 562 size_t count, loff_t *pos) 563 { 564 struct ib_uverbs_file *file = filp->private_data; 565 const struct uverbs_api_write_method *method_elm; 566 struct uverbs_api *uapi = file->device->uapi; 567 struct ib_uverbs_ex_cmd_hdr ex_hdr; 568 struct ib_uverbs_cmd_hdr hdr; 569 struct uverbs_attr_bundle bundle; 570 int srcu_key; 571 ssize_t ret; 572 573 if (!ib_safe_file_access(filp)) { 574 pr_err_once("uverbs_write: process %d (%s) changed security contexts after opening file descriptor, this is not allowed.\n", 575 task_tgid_vnr(current), current->comm); 576 return -EACCES; 577 } 578 579 if (count < sizeof(hdr)) 580 return -EINVAL; 581 582 if (copy_from_user(&hdr, buf, sizeof(hdr))) 583 return -EFAULT; 584 585 method_elm = uapi_get_method(uapi, hdr.command); 586 if (IS_ERR(method_elm)) 587 return PTR_ERR(method_elm); 588 589 if (method_elm->is_ex) { 590 if (count < (sizeof(hdr) + sizeof(ex_hdr))) 591 return -EINVAL; 592 if (copy_from_user(&ex_hdr, buf + sizeof(hdr), sizeof(ex_hdr))) 593 return -EFAULT; 594 } 595 596 ret = verify_hdr(&hdr, &ex_hdr, count, method_elm); 597 if (ret) 598 return ret; 599 600 srcu_key = srcu_read_lock(&file->device->disassociate_srcu); 601 602 buf += sizeof(hdr); 603 604 memset(bundle.attr_present, 0, sizeof(bundle.attr_present)); 605 bundle.ufile = file; 606 bundle.context = NULL; /* only valid if bundle has uobject */ 607 bundle.uobject = NULL; 608 if (!method_elm->is_ex) { 609 size_t in_len = hdr.in_words * 4 - sizeof(hdr); 610 size_t out_len = hdr.out_words * 4; 611 u64 response = 0; 612 613 if (method_elm->has_udata) { 614 bundle.driver_udata.inlen = 615 in_len - method_elm->req_size; 616 in_len = method_elm->req_size; 617 if (bundle.driver_udata.inlen) 618 bundle.driver_udata.inbuf = buf + in_len; 619 else 620 bundle.driver_udata.inbuf = NULL; 621 } else { 622 memset(&bundle.driver_udata, 0, 623 sizeof(bundle.driver_udata)); 624 } 625 626 if (method_elm->has_resp) { 627 /* 628 * The macros check that if has_resp is set 629 * then the command request structure starts 630 * with a '__aligned u64 response' member. 631 */ 632 ret = get_user(response, (const u64 __user *)buf); 633 if (ret) 634 goto out_unlock; 635 636 if (method_elm->has_udata) { 637 bundle.driver_udata.outlen = 638 out_len - method_elm->resp_size; 639 out_len = method_elm->resp_size; 640 if (bundle.driver_udata.outlen) 641 bundle.driver_udata.outbuf = 642 u64_to_user_ptr(response + 643 out_len); 644 else 645 bundle.driver_udata.outbuf = NULL; 646 } 647 } else { 648 bundle.driver_udata.outlen = 0; 649 bundle.driver_udata.outbuf = NULL; 650 } 651 652 ib_uverbs_init_udata_buf_or_null( 653 &bundle.ucore, buf, u64_to_user_ptr(response), 654 in_len, out_len); 655 } else { 656 buf += sizeof(ex_hdr); 657 658 ib_uverbs_init_udata_buf_or_null(&bundle.ucore, buf, 659 u64_to_user_ptr(ex_hdr.response), 660 hdr.in_words * 8, hdr.out_words * 8); 661 662 ib_uverbs_init_udata_buf_or_null( 663 &bundle.driver_udata, buf + bundle.ucore.inlen, 664 u64_to_user_ptr(ex_hdr.response) + bundle.ucore.outlen, 665 ex_hdr.provider_in_words * 8, 666 ex_hdr.provider_out_words * 8); 667 668 } 669 670 ret = method_elm->handler(&bundle); 671 if (bundle.uobject) 672 uverbs_finalize_object(bundle.uobject, UVERBS_ACCESS_NEW, true, 673 !ret, &bundle); 674 out_unlock: 675 srcu_read_unlock(&file->device->disassociate_srcu, srcu_key); 676 return (ret) ? : count; 677 } 678 679 static const struct vm_operations_struct rdma_umap_ops; 680 681 static int ib_uverbs_mmap(struct file *filp, struct vm_area_struct *vma) 682 { 683 struct ib_uverbs_file *file = filp->private_data; 684 struct ib_ucontext *ucontext; 685 int ret = 0; 686 int srcu_key; 687 688 srcu_key = srcu_read_lock(&file->device->disassociate_srcu); 689 ucontext = ib_uverbs_get_ucontext_file(file); 690 if (IS_ERR(ucontext)) { 691 ret = PTR_ERR(ucontext); 692 goto out; 693 } 694 vma->vm_ops = &rdma_umap_ops; 695 ret = ucontext->device->ops.mmap(ucontext, vma); 696 out: 697 srcu_read_unlock(&file->device->disassociate_srcu, srcu_key); 698 return ret; 699 } 700 701 /* 702 * The VMA has been dup'd, initialize the vm_private_data with a new tracking 703 * struct 704 */ 705 static void rdma_umap_open(struct vm_area_struct *vma) 706 { 707 struct ib_uverbs_file *ufile = vma->vm_file->private_data; 708 struct rdma_umap_priv *opriv = vma->vm_private_data; 709 struct rdma_umap_priv *priv; 710 711 if (!opriv) 712 return; 713 714 /* We are racing with disassociation */ 715 if (!down_read_trylock(&ufile->hw_destroy_rwsem)) 716 goto out_zap; 717 /* 718 * Disassociation already completed, the VMA should already be zapped. 719 */ 720 if (!ufile->ucontext) 721 goto out_unlock; 722 723 priv = kzalloc(sizeof(*priv), GFP_KERNEL); 724 if (!priv) 725 goto out_unlock; 726 rdma_umap_priv_init(priv, vma, opriv->entry); 727 728 up_read(&ufile->hw_destroy_rwsem); 729 return; 730 731 out_unlock: 732 up_read(&ufile->hw_destroy_rwsem); 733 out_zap: 734 /* 735 * We can't allow the VMA to be created with the actual IO pages, that 736 * would break our API contract, and it can't be stopped at this 737 * point, so zap it. 738 */ 739 vma->vm_private_data = NULL; 740 zap_vma_ptes(vma, vma->vm_start, vma->vm_end - vma->vm_start); 741 } 742 743 static void rdma_umap_close(struct vm_area_struct *vma) 744 { 745 struct ib_uverbs_file *ufile = vma->vm_file->private_data; 746 struct rdma_umap_priv *priv = vma->vm_private_data; 747 748 if (!priv) 749 return; 750 751 /* 752 * The vma holds a reference on the struct file that created it, which 753 * in turn means that the ib_uverbs_file is guaranteed to exist at 754 * this point. 755 */ 756 mutex_lock(&ufile->umap_lock); 757 if (priv->entry) 758 rdma_user_mmap_entry_put(priv->entry); 759 760 list_del(&priv->list); 761 mutex_unlock(&ufile->umap_lock); 762 kfree(priv); 763 } 764 765 /* 766 * Once the zap_vma_ptes has been called touches to the VMA will come here and 767 * we return a dummy writable zero page for all the pfns. 768 */ 769 static vm_fault_t rdma_umap_fault(struct vm_fault *vmf) 770 { 771 struct ib_uverbs_file *ufile = vmf->vma->vm_file->private_data; 772 struct rdma_umap_priv *priv = vmf->vma->vm_private_data; 773 vm_fault_t ret = 0; 774 775 if (!priv) 776 return VM_FAULT_SIGBUS; 777 778 /* Read only pages can just use the system zero page. */ 779 if (!(vmf->vma->vm_flags & (VM_WRITE | VM_MAYWRITE))) { 780 vmf->page = ZERO_PAGE(vmf->address); 781 get_page(vmf->page); 782 return 0; 783 } 784 785 mutex_lock(&ufile->umap_lock); 786 if (!ufile->disassociate_page) 787 ufile->disassociate_page = 788 alloc_pages(vmf->gfp_mask | __GFP_ZERO, 0); 789 790 if (ufile->disassociate_page) { 791 /* 792 * This VMA is forced to always be shared so this doesn't have 793 * to worry about COW. 794 */ 795 vmf->page = ufile->disassociate_page; 796 get_page(vmf->page); 797 } else { 798 ret = VM_FAULT_SIGBUS; 799 } 800 mutex_unlock(&ufile->umap_lock); 801 802 return ret; 803 } 804 805 static const struct vm_operations_struct rdma_umap_ops = { 806 .open = rdma_umap_open, 807 .close = rdma_umap_close, 808 .fault = rdma_umap_fault, 809 }; 810 811 void uverbs_user_mmap_disassociate(struct ib_uverbs_file *ufile) 812 { 813 struct rdma_umap_priv *priv, *next_priv; 814 815 lockdep_assert_held(&ufile->hw_destroy_rwsem); 816 817 while (1) { 818 struct mm_struct *mm = NULL; 819 820 /* Get an arbitrary mm pointer that hasn't been cleaned yet */ 821 mutex_lock(&ufile->umap_lock); 822 while (!list_empty(&ufile->umaps)) { 823 int ret; 824 825 priv = list_first_entry(&ufile->umaps, 826 struct rdma_umap_priv, list); 827 mm = priv->vma->vm_mm; 828 ret = mmget_not_zero(mm); 829 if (!ret) { 830 list_del_init(&priv->list); 831 if (priv->entry) { 832 rdma_user_mmap_entry_put(priv->entry); 833 priv->entry = NULL; 834 } 835 mm = NULL; 836 continue; 837 } 838 break; 839 } 840 mutex_unlock(&ufile->umap_lock); 841 if (!mm) 842 return; 843 844 /* 845 * The umap_lock is nested under mmap_lock since it used within 846 * the vma_ops callbacks, so we have to clean the list one mm 847 * at a time to get the lock ordering right. Typically there 848 * will only be one mm, so no big deal. 849 */ 850 mmap_read_lock(mm); 851 mutex_lock(&ufile->umap_lock); 852 list_for_each_entry_safe (priv, next_priv, &ufile->umaps, 853 list) { 854 struct vm_area_struct *vma = priv->vma; 855 856 if (vma->vm_mm != mm) 857 continue; 858 list_del_init(&priv->list); 859 860 zap_vma_ptes(vma, vma->vm_start, 861 vma->vm_end - vma->vm_start); 862 863 if (priv->entry) { 864 rdma_user_mmap_entry_put(priv->entry); 865 priv->entry = NULL; 866 } 867 } 868 mutex_unlock(&ufile->umap_lock); 869 mmap_read_unlock(mm); 870 mmput(mm); 871 } 872 } 873 874 /* 875 * ib_uverbs_open() does not need the BKL: 876 * 877 * - the ib_uverbs_device structures are properly reference counted and 878 * everything else is purely local to the file being created, so 879 * races against other open calls are not a problem; 880 * - there is no ioctl method to race against; 881 * - the open method will either immediately run -ENXIO, or all 882 * required initialization will be done. 883 */ 884 static int ib_uverbs_open(struct inode *inode, struct file *filp) 885 { 886 struct ib_uverbs_device *dev; 887 struct ib_uverbs_file *file; 888 struct ib_device *ib_dev; 889 int ret; 890 int module_dependent; 891 int srcu_key; 892 893 dev = container_of(inode->i_cdev, struct ib_uverbs_device, cdev); 894 if (!atomic_inc_not_zero(&dev->refcount)) 895 return -ENXIO; 896 897 get_device(&dev->dev); 898 srcu_key = srcu_read_lock(&dev->disassociate_srcu); 899 mutex_lock(&dev->lists_mutex); 900 ib_dev = srcu_dereference(dev->ib_dev, 901 &dev->disassociate_srcu); 902 if (!ib_dev) { 903 ret = -EIO; 904 goto err; 905 } 906 907 if (!rdma_dev_access_netns(ib_dev, current->nsproxy->net_ns)) { 908 ret = -EPERM; 909 goto err; 910 } 911 912 /* In case IB device supports disassociate ucontext, there is no hard 913 * dependency between uverbs device and its low level device. 914 */ 915 module_dependent = !(ib_dev->ops.disassociate_ucontext); 916 917 if (module_dependent) { 918 if (!try_module_get(ib_dev->ops.owner)) { 919 ret = -ENODEV; 920 goto err; 921 } 922 } 923 924 file = kzalloc(sizeof(*file), GFP_KERNEL); 925 if (!file) { 926 ret = -ENOMEM; 927 if (module_dependent) 928 goto err_module; 929 930 goto err; 931 } 932 933 file->device = dev; 934 kref_init(&file->ref); 935 mutex_init(&file->ucontext_lock); 936 937 spin_lock_init(&file->uobjects_lock); 938 INIT_LIST_HEAD(&file->uobjects); 939 init_rwsem(&file->hw_destroy_rwsem); 940 mutex_init(&file->umap_lock); 941 INIT_LIST_HEAD(&file->umaps); 942 943 filp->private_data = file; 944 list_add_tail(&file->list, &dev->uverbs_file_list); 945 mutex_unlock(&dev->lists_mutex); 946 srcu_read_unlock(&dev->disassociate_srcu, srcu_key); 947 948 setup_ufile_idr_uobject(file); 949 950 return stream_open(inode, filp); 951 952 err_module: 953 module_put(ib_dev->ops.owner); 954 955 err: 956 mutex_unlock(&dev->lists_mutex); 957 srcu_read_unlock(&dev->disassociate_srcu, srcu_key); 958 if (atomic_dec_and_test(&dev->refcount)) 959 ib_uverbs_comp_dev(dev); 960 961 put_device(&dev->dev); 962 return ret; 963 } 964 965 static int ib_uverbs_close(struct inode *inode, struct file *filp) 966 { 967 struct ib_uverbs_file *file = filp->private_data; 968 969 uverbs_destroy_ufile_hw(file, RDMA_REMOVE_CLOSE); 970 971 mutex_lock(&file->device->lists_mutex); 972 list_del_init(&file->list); 973 mutex_unlock(&file->device->lists_mutex); 974 975 kref_put(&file->ref, ib_uverbs_release_file); 976 977 return 0; 978 } 979 980 static const struct file_operations uverbs_fops = { 981 .owner = THIS_MODULE, 982 .write = ib_uverbs_write, 983 .open = ib_uverbs_open, 984 .release = ib_uverbs_close, 985 .llseek = no_llseek, 986 .unlocked_ioctl = ib_uverbs_ioctl, 987 .compat_ioctl = compat_ptr_ioctl, 988 }; 989 990 static const struct file_operations uverbs_mmap_fops = { 991 .owner = THIS_MODULE, 992 .write = ib_uverbs_write, 993 .mmap = ib_uverbs_mmap, 994 .open = ib_uverbs_open, 995 .release = ib_uverbs_close, 996 .llseek = no_llseek, 997 .unlocked_ioctl = ib_uverbs_ioctl, 998 .compat_ioctl = compat_ptr_ioctl, 999 }; 1000 1001 static int ib_uverbs_get_nl_info(struct ib_device *ibdev, void *client_data, 1002 struct ib_client_nl_info *res) 1003 { 1004 struct ib_uverbs_device *uverbs_dev = client_data; 1005 int ret; 1006 1007 if (res->port != -1) 1008 return -EINVAL; 1009 1010 res->abi = ibdev->ops.uverbs_abi_ver; 1011 res->cdev = &uverbs_dev->dev; 1012 1013 /* 1014 * To support DRIVER_ID binding in userspace some of the driver need 1015 * upgrading to expose their PCI dependent revision information 1016 * through get_context instead of relying on modalias matching. When 1017 * the drivers are fixed they can drop this flag. 1018 */ 1019 if (!ibdev->ops.uverbs_no_driver_id_binding) { 1020 ret = nla_put_u32(res->nl_msg, RDMA_NLDEV_ATTR_UVERBS_DRIVER_ID, 1021 ibdev->ops.driver_id); 1022 if (ret) 1023 return ret; 1024 } 1025 return 0; 1026 } 1027 1028 static struct ib_client uverbs_client = { 1029 .name = "uverbs", 1030 .no_kverbs_req = true, 1031 .add = ib_uverbs_add_one, 1032 .remove = ib_uverbs_remove_one, 1033 .get_nl_info = ib_uverbs_get_nl_info, 1034 }; 1035 MODULE_ALIAS_RDMA_CLIENT("uverbs"); 1036 1037 static ssize_t ibdev_show(struct device *device, struct device_attribute *attr, 1038 char *buf) 1039 { 1040 struct ib_uverbs_device *dev = 1041 container_of(device, struct ib_uverbs_device, dev); 1042 int ret = -ENODEV; 1043 int srcu_key; 1044 struct ib_device *ib_dev; 1045 1046 srcu_key = srcu_read_lock(&dev->disassociate_srcu); 1047 ib_dev = srcu_dereference(dev->ib_dev, &dev->disassociate_srcu); 1048 if (ib_dev) 1049 ret = sysfs_emit(buf, "%s\n", dev_name(&ib_dev->dev)); 1050 srcu_read_unlock(&dev->disassociate_srcu, srcu_key); 1051 1052 return ret; 1053 } 1054 static DEVICE_ATTR_RO(ibdev); 1055 1056 static ssize_t abi_version_show(struct device *device, 1057 struct device_attribute *attr, char *buf) 1058 { 1059 struct ib_uverbs_device *dev = 1060 container_of(device, struct ib_uverbs_device, dev); 1061 int ret = -ENODEV; 1062 int srcu_key; 1063 struct ib_device *ib_dev; 1064 1065 srcu_key = srcu_read_lock(&dev->disassociate_srcu); 1066 ib_dev = srcu_dereference(dev->ib_dev, &dev->disassociate_srcu); 1067 if (ib_dev) 1068 ret = sysfs_emit(buf, "%u\n", ib_dev->ops.uverbs_abi_ver); 1069 srcu_read_unlock(&dev->disassociate_srcu, srcu_key); 1070 1071 return ret; 1072 } 1073 static DEVICE_ATTR_RO(abi_version); 1074 1075 static struct attribute *ib_dev_attrs[] = { 1076 &dev_attr_abi_version.attr, 1077 &dev_attr_ibdev.attr, 1078 NULL, 1079 }; 1080 1081 static const struct attribute_group dev_attr_group = { 1082 .attrs = ib_dev_attrs, 1083 }; 1084 1085 static CLASS_ATTR_STRING(abi_version, S_IRUGO, 1086 __stringify(IB_USER_VERBS_ABI_VERSION)); 1087 1088 static int ib_uverbs_create_uapi(struct ib_device *device, 1089 struct ib_uverbs_device *uverbs_dev) 1090 { 1091 struct uverbs_api *uapi; 1092 1093 uapi = uverbs_alloc_api(device); 1094 if (IS_ERR(uapi)) 1095 return PTR_ERR(uapi); 1096 1097 uverbs_dev->uapi = uapi; 1098 return 0; 1099 } 1100 1101 static int ib_uverbs_add_one(struct ib_device *device) 1102 { 1103 int devnum; 1104 dev_t base; 1105 struct ib_uverbs_device *uverbs_dev; 1106 int ret; 1107 1108 if (!device->ops.alloc_ucontext) 1109 return -EOPNOTSUPP; 1110 1111 uverbs_dev = kzalloc(sizeof(*uverbs_dev), GFP_KERNEL); 1112 if (!uverbs_dev) 1113 return -ENOMEM; 1114 1115 ret = init_srcu_struct(&uverbs_dev->disassociate_srcu); 1116 if (ret) { 1117 kfree(uverbs_dev); 1118 return -ENOMEM; 1119 } 1120 1121 device_initialize(&uverbs_dev->dev); 1122 uverbs_dev->dev.class = uverbs_class; 1123 uverbs_dev->dev.parent = device->dev.parent; 1124 uverbs_dev->dev.release = ib_uverbs_release_dev; 1125 uverbs_dev->groups[0] = &dev_attr_group; 1126 uverbs_dev->dev.groups = uverbs_dev->groups; 1127 atomic_set(&uverbs_dev->refcount, 1); 1128 init_completion(&uverbs_dev->comp); 1129 uverbs_dev->xrcd_tree = RB_ROOT; 1130 mutex_init(&uverbs_dev->xrcd_tree_mutex); 1131 mutex_init(&uverbs_dev->lists_mutex); 1132 INIT_LIST_HEAD(&uverbs_dev->uverbs_file_list); 1133 rcu_assign_pointer(uverbs_dev->ib_dev, device); 1134 uverbs_dev->num_comp_vectors = device->num_comp_vectors; 1135 1136 devnum = ida_alloc_max(&uverbs_ida, IB_UVERBS_MAX_DEVICES - 1, 1137 GFP_KERNEL); 1138 if (devnum < 0) { 1139 ret = -ENOMEM; 1140 goto err; 1141 } 1142 uverbs_dev->devnum = devnum; 1143 if (devnum >= IB_UVERBS_NUM_FIXED_MINOR) 1144 base = dynamic_uverbs_dev + devnum - IB_UVERBS_NUM_FIXED_MINOR; 1145 else 1146 base = IB_UVERBS_BASE_DEV + devnum; 1147 1148 ret = ib_uverbs_create_uapi(device, uverbs_dev); 1149 if (ret) 1150 goto err_uapi; 1151 1152 uverbs_dev->dev.devt = base; 1153 dev_set_name(&uverbs_dev->dev, "uverbs%d", uverbs_dev->devnum); 1154 1155 cdev_init(&uverbs_dev->cdev, 1156 device->ops.mmap ? &uverbs_mmap_fops : &uverbs_fops); 1157 uverbs_dev->cdev.owner = THIS_MODULE; 1158 1159 ret = cdev_device_add(&uverbs_dev->cdev, &uverbs_dev->dev); 1160 if (ret) 1161 goto err_uapi; 1162 1163 ib_set_client_data(device, &uverbs_client, uverbs_dev); 1164 return 0; 1165 1166 err_uapi: 1167 ida_free(&uverbs_ida, devnum); 1168 err: 1169 if (atomic_dec_and_test(&uverbs_dev->refcount)) 1170 ib_uverbs_comp_dev(uverbs_dev); 1171 wait_for_completion(&uverbs_dev->comp); 1172 put_device(&uverbs_dev->dev); 1173 return ret; 1174 } 1175 1176 static void ib_uverbs_free_hw_resources(struct ib_uverbs_device *uverbs_dev, 1177 struct ib_device *ib_dev) 1178 { 1179 struct ib_uverbs_file *file; 1180 1181 /* Pending running commands to terminate */ 1182 uverbs_disassociate_api_pre(uverbs_dev); 1183 1184 mutex_lock(&uverbs_dev->lists_mutex); 1185 while (!list_empty(&uverbs_dev->uverbs_file_list)) { 1186 file = list_first_entry(&uverbs_dev->uverbs_file_list, 1187 struct ib_uverbs_file, list); 1188 list_del_init(&file->list); 1189 kref_get(&file->ref); 1190 1191 /* We must release the mutex before going ahead and calling 1192 * uverbs_cleanup_ufile, as it might end up indirectly calling 1193 * uverbs_close, for example due to freeing the resources (e.g 1194 * mmput). 1195 */ 1196 mutex_unlock(&uverbs_dev->lists_mutex); 1197 1198 uverbs_destroy_ufile_hw(file, RDMA_REMOVE_DRIVER_REMOVE); 1199 kref_put(&file->ref, ib_uverbs_release_file); 1200 1201 mutex_lock(&uverbs_dev->lists_mutex); 1202 } 1203 mutex_unlock(&uverbs_dev->lists_mutex); 1204 1205 uverbs_disassociate_api(uverbs_dev->uapi); 1206 } 1207 1208 static void ib_uverbs_remove_one(struct ib_device *device, void *client_data) 1209 { 1210 struct ib_uverbs_device *uverbs_dev = client_data; 1211 int wait_clients = 1; 1212 1213 cdev_device_del(&uverbs_dev->cdev, &uverbs_dev->dev); 1214 ida_free(&uverbs_ida, uverbs_dev->devnum); 1215 1216 if (device->ops.disassociate_ucontext) { 1217 /* We disassociate HW resources and immediately return. 1218 * Userspace will see a EIO errno for all future access. 1219 * Upon returning, ib_device may be freed internally and is not 1220 * valid any more. 1221 * uverbs_device is still available until all clients close 1222 * their files, then the uverbs device ref count will be zero 1223 * and its resources will be freed. 1224 * Note: At this point no more files can be opened since the 1225 * cdev was deleted, however active clients can still issue 1226 * commands and close their open files. 1227 */ 1228 ib_uverbs_free_hw_resources(uverbs_dev, device); 1229 wait_clients = 0; 1230 } 1231 1232 if (atomic_dec_and_test(&uverbs_dev->refcount)) 1233 ib_uverbs_comp_dev(uverbs_dev); 1234 if (wait_clients) 1235 wait_for_completion(&uverbs_dev->comp); 1236 1237 put_device(&uverbs_dev->dev); 1238 } 1239 1240 static char *uverbs_devnode(struct device *dev, umode_t *mode) 1241 { 1242 if (mode) 1243 *mode = 0666; 1244 return kasprintf(GFP_KERNEL, "infiniband/%s", dev_name(dev)); 1245 } 1246 1247 static int __init ib_uverbs_init(void) 1248 { 1249 int ret; 1250 1251 ret = register_chrdev_region(IB_UVERBS_BASE_DEV, 1252 IB_UVERBS_NUM_FIXED_MINOR, 1253 "infiniband_verbs"); 1254 if (ret) { 1255 pr_err("user_verbs: couldn't register device number\n"); 1256 goto out; 1257 } 1258 1259 ret = alloc_chrdev_region(&dynamic_uverbs_dev, 0, 1260 IB_UVERBS_NUM_DYNAMIC_MINOR, 1261 "infiniband_verbs"); 1262 if (ret) { 1263 pr_err("couldn't register dynamic device number\n"); 1264 goto out_alloc; 1265 } 1266 1267 uverbs_class = class_create(THIS_MODULE, "infiniband_verbs"); 1268 if (IS_ERR(uverbs_class)) { 1269 ret = PTR_ERR(uverbs_class); 1270 pr_err("user_verbs: couldn't create class infiniband_verbs\n"); 1271 goto out_chrdev; 1272 } 1273 1274 uverbs_class->devnode = uverbs_devnode; 1275 1276 ret = class_create_file(uverbs_class, &class_attr_abi_version.attr); 1277 if (ret) { 1278 pr_err("user_verbs: couldn't create abi_version attribute\n"); 1279 goto out_class; 1280 } 1281 1282 ret = ib_register_client(&uverbs_client); 1283 if (ret) { 1284 pr_err("user_verbs: couldn't register client\n"); 1285 goto out_class; 1286 } 1287 1288 return 0; 1289 1290 out_class: 1291 class_destroy(uverbs_class); 1292 1293 out_chrdev: 1294 unregister_chrdev_region(dynamic_uverbs_dev, 1295 IB_UVERBS_NUM_DYNAMIC_MINOR); 1296 1297 out_alloc: 1298 unregister_chrdev_region(IB_UVERBS_BASE_DEV, 1299 IB_UVERBS_NUM_FIXED_MINOR); 1300 1301 out: 1302 return ret; 1303 } 1304 1305 static void __exit ib_uverbs_cleanup(void) 1306 { 1307 ib_unregister_client(&uverbs_client); 1308 class_destroy(uverbs_class); 1309 unregister_chrdev_region(IB_UVERBS_BASE_DEV, 1310 IB_UVERBS_NUM_FIXED_MINOR); 1311 unregister_chrdev_region(dynamic_uverbs_dev, 1312 IB_UVERBS_NUM_DYNAMIC_MINOR); 1313 mmu_notifier_synchronize(); 1314 } 1315 1316 module_init(ib_uverbs_init); 1317 module_exit(ib_uverbs_cleanup); 1318