Linux Audio

Check our new training course

Loading...
v6.8
   1/*
   2 * Copyright (c) 2005-2006 Intel Corporation.  All rights reserved.
   3 *
   4 * This software is available to you under a choice of one of two
   5 * licenses.  You may choose to be licensed under the terms of the GNU
   6 * General Public License (GPL) Version 2, available from the file
   7 * COPYING in the main directory of this source tree, or the
   8 * OpenIB.org BSD license below:
   9 *
  10 *     Redistribution and use in source and binary forms, with or
  11 *     without modification, are permitted provided that the following
  12 *     conditions are met:
  13 *
  14 *      - Redistributions of source code must retain the above
  15 *	copyright notice, this list of conditions and the following
  16 *	disclaimer.
  17 *
  18 *      - Redistributions in binary form must reproduce the above
  19 *	copyright notice, this list of conditions and the following
  20 *	disclaimer in the documentation and/or other materials
  21 *	provided with the distribution.
  22 *
  23 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
  24 * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
  25 * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
  26 * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
  27 * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
  28 * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
  29 * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
  30 * SOFTWARE.
  31 */
  32
  33#include <linux/completion.h>
  34#include <linux/file.h>
  35#include <linux/mutex.h>
  36#include <linux/poll.h>
  37#include <linux/sched.h>
  38#include <linux/idr.h>
  39#include <linux/in.h>
  40#include <linux/in6.h>
  41#include <linux/miscdevice.h>
  42#include <linux/slab.h>
  43#include <linux/sysctl.h>
  44#include <linux/module.h>
  45#include <linux/nsproxy.h>
  46
  47#include <linux/nospec.h>
  48
  49#include <rdma/rdma_user_cm.h>
  50#include <rdma/ib_marshall.h>
  51#include <rdma/rdma_cm.h>
  52#include <rdma/rdma_cm_ib.h>
  53#include <rdma/ib_addr.h>
  54#include <rdma/ib.h>
  55#include <rdma/ib_cm.h>
  56#include <rdma/rdma_netlink.h>
  57#include "core_priv.h"
  58
  59MODULE_AUTHOR("Sean Hefty");
  60MODULE_DESCRIPTION("RDMA Userspace Connection Manager Access");
  61MODULE_LICENSE("Dual BSD/GPL");
  62
  63static unsigned int max_backlog = 1024;
  64
  65static struct ctl_table_header *ucma_ctl_table_hdr;
  66static struct ctl_table ucma_ctl_table[] = {
  67	{
  68		.procname	= "max_backlog",
  69		.data		= &max_backlog,
  70		.maxlen		= sizeof max_backlog,
  71		.mode		= 0644,
  72		.proc_handler	= proc_dointvec,
  73	},
 
  74};
  75
  76struct ucma_file {
  77	struct mutex		mut;
  78	struct file		*filp;
  79	struct list_head	ctx_list;
  80	struct list_head	event_list;
  81	wait_queue_head_t	poll_wait;
 
  82};
  83
  84struct ucma_context {
  85	u32			id;
  86	struct completion	comp;
  87	refcount_t		ref;
  88	int			events_reported;
  89	atomic_t		backlog;
  90
  91	struct ucma_file	*file;
  92	struct rdma_cm_id	*cm_id;
  93	struct mutex		mutex;
  94	u64			uid;
  95
  96	struct list_head	list;
  97	struct list_head	mc_list;
 
 
 
 
 
 
  98	struct work_struct	close_work;
  99};
 100
 101struct ucma_multicast {
 102	struct ucma_context	*ctx;
 103	u32			id;
 104	int			events_reported;
 105
 106	u64			uid;
 107	u8			join_state;
 108	struct list_head	list;
 109	struct sockaddr_storage	addr;
 110};
 111
 112struct ucma_event {
 113	struct ucma_context	*ctx;
 114	struct ucma_context	*conn_req_ctx;
 115	struct ucma_multicast	*mc;
 116	struct list_head	list;
 
 117	struct rdma_ucm_event_resp resp;
 
 118};
 119
 120static DEFINE_XARRAY_ALLOC(ctx_table);
 121static DEFINE_XARRAY_ALLOC(multicast_table);
 122
 123static const struct file_operations ucma_fops;
 124static int ucma_destroy_private_ctx(struct ucma_context *ctx);
 125
 126static inline struct ucma_context *_ucma_find_context(int id,
 127						      struct ucma_file *file)
 128{
 129	struct ucma_context *ctx;
 130
 131	ctx = xa_load(&ctx_table, id);
 132	if (!ctx)
 133		ctx = ERR_PTR(-ENOENT);
 134	else if (ctx->file != file)
 135		ctx = ERR_PTR(-EINVAL);
 136	return ctx;
 137}
 138
 139static struct ucma_context *ucma_get_ctx(struct ucma_file *file, int id)
 140{
 141	struct ucma_context *ctx;
 142
 143	xa_lock(&ctx_table);
 144	ctx = _ucma_find_context(id, file);
 145	if (!IS_ERR(ctx))
 146		if (!refcount_inc_not_zero(&ctx->ref))
 147			ctx = ERR_PTR(-ENXIO);
 148	xa_unlock(&ctx_table);
 
 
 
 149	return ctx;
 150}
 151
 152static void ucma_put_ctx(struct ucma_context *ctx)
 153{
 154	if (refcount_dec_and_test(&ctx->ref))
 155		complete(&ctx->comp);
 156}
 157
 158/*
 159 * Same as ucm_get_ctx but requires that ->cm_id->device is valid, eg that the
 160 * CM_ID is bound.
 161 */
 162static struct ucma_context *ucma_get_ctx_dev(struct ucma_file *file, int id)
 163{
 164	struct ucma_context *ctx = ucma_get_ctx(file, id);
 165
 166	if (IS_ERR(ctx))
 167		return ctx;
 168	if (!ctx->cm_id->device) {
 169		ucma_put_ctx(ctx);
 170		return ERR_PTR(-EINVAL);
 171	}
 172	return ctx;
 173}
 174
 
 
 
 
 
 
 
 
 175static void ucma_close_id(struct work_struct *work)
 176{
 177	struct ucma_context *ctx =  container_of(work, struct ucma_context, close_work);
 178
 179	/* once all inflight tasks are finished, we close all underlying
 180	 * resources. The context is still alive till its explicit destryoing
 181	 * by its creator. This puts back the xarray's reference.
 182	 */
 183	ucma_put_ctx(ctx);
 184	wait_for_completion(&ctx->comp);
 185	/* No new events will be generated after destroying the id. */
 186	rdma_destroy_id(ctx->cm_id);
 187
 188	/* Reading the cm_id without holding a positive ref is not allowed */
 189	ctx->cm_id = NULL;
 190}
 191
 192static struct ucma_context *ucma_alloc_ctx(struct ucma_file *file)
 193{
 194	struct ucma_context *ctx;
 195
 196	ctx = kzalloc(sizeof(*ctx), GFP_KERNEL);
 197	if (!ctx)
 198		return NULL;
 199
 200	INIT_WORK(&ctx->close_work, ucma_close_id);
 
 201	init_completion(&ctx->comp);
 202	INIT_LIST_HEAD(&ctx->mc_list);
 203	/* So list_del() will work if we don't do ucma_finish_ctx() */
 204	INIT_LIST_HEAD(&ctx->list);
 205	ctx->file = file;
 206	mutex_init(&ctx->mutex);
 207
 208	if (xa_alloc(&ctx_table, &ctx->id, NULL, xa_limit_32b, GFP_KERNEL)) {
 209		kfree(ctx);
 210		return NULL;
 211	}
 
 
 
 212	return ctx;
 213}
 214
 215static void ucma_set_ctx_cm_id(struct ucma_context *ctx,
 216			       struct rdma_cm_id *cm_id)
 217{
 218	refcount_set(&ctx->ref, 1);
 219	ctx->cm_id = cm_id;
 220}
 221
 222static void ucma_finish_ctx(struct ucma_context *ctx)
 223{
 224	lockdep_assert_held(&ctx->file->mut);
 225	list_add_tail(&ctx->list, &ctx->file->ctx_list);
 226	xa_store(&ctx_table, ctx->id, ctx, GFP_KERNEL);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 227}
 228
 229static void ucma_copy_conn_event(struct rdma_ucm_conn_param *dst,
 230				 struct rdma_conn_param *src)
 231{
 232	if (src->private_data_len)
 233		memcpy(dst->private_data, src->private_data,
 234		       src->private_data_len);
 235	dst->private_data_len = src->private_data_len;
 236	dst->responder_resources = src->responder_resources;
 237	dst->initiator_depth = src->initiator_depth;
 238	dst->flow_control = src->flow_control;
 239	dst->retry_count = src->retry_count;
 240	dst->rnr_retry_count = src->rnr_retry_count;
 241	dst->srq = src->srq;
 242	dst->qp_num = src->qp_num;
 243}
 244
 245static void ucma_copy_ud_event(struct ib_device *device,
 246			       struct rdma_ucm_ud_param *dst,
 247			       struct rdma_ud_param *src)
 248{
 249	if (src->private_data_len)
 250		memcpy(dst->private_data, src->private_data,
 251		       src->private_data_len);
 252	dst->private_data_len = src->private_data_len;
 253	ib_copy_ah_attr_to_user(device, &dst->ah_attr, &src->ah_attr);
 254	dst->qp_num = src->qp_num;
 255	dst->qkey = src->qkey;
 256}
 257
 258static struct ucma_event *ucma_create_uevent(struct ucma_context *ctx,
 259					     struct rdma_cm_event *event)
 
 260{
 261	struct ucma_event *uevent;
 262
 263	uevent = kzalloc(sizeof(*uevent), GFP_KERNEL);
 264	if (!uevent)
 265		return NULL;
 266
 267	uevent->ctx = ctx;
 268	switch (event->event) {
 269	case RDMA_CM_EVENT_MULTICAST_JOIN:
 270	case RDMA_CM_EVENT_MULTICAST_ERROR:
 271		uevent->mc = (struct ucma_multicast *)
 272			     event->param.ud.private_data;
 273		uevent->resp.uid = uevent->mc->uid;
 274		uevent->resp.id = uevent->mc->id;
 275		break;
 276	default:
 277		uevent->resp.uid = ctx->uid;
 278		uevent->resp.id = ctx->id;
 279		break;
 280	}
 281	uevent->resp.event = event->event;
 282	uevent->resp.status = event->status;
 283	if (ctx->cm_id->qp_type == IB_QPT_UD)
 284		ucma_copy_ud_event(ctx->cm_id->device, &uevent->resp.param.ud,
 285				   &event->param.ud);
 286	else
 287		ucma_copy_conn_event(&uevent->resp.param.conn,
 288				     &event->param.conn);
 289
 290	uevent->resp.ece.vendor_id = event->ece.vendor_id;
 291	uevent->resp.ece.attr_mod = event->ece.attr_mod;
 292	return uevent;
 293}
 294
 295static int ucma_connect_event_handler(struct rdma_cm_id *cm_id,
 296				      struct rdma_cm_event *event)
 297{
 298	struct ucma_context *listen_ctx = cm_id->context;
 299	struct ucma_context *ctx;
 300	struct ucma_event *uevent;
 301
 302	if (!atomic_add_unless(&listen_ctx->backlog, -1, 0))
 303		return -ENOMEM;
 304	ctx = ucma_alloc_ctx(listen_ctx->file);
 305	if (!ctx)
 306		goto err_backlog;
 307	ucma_set_ctx_cm_id(ctx, cm_id);
 308
 309	uevent = ucma_create_uevent(listen_ctx, event);
 310	if (!uevent)
 311		goto err_alloc;
 312	uevent->conn_req_ctx = ctx;
 313	uevent->resp.id = ctx->id;
 314
 315	ctx->cm_id->context = ctx;
 
 316
 317	mutex_lock(&ctx->file->mut);
 318	ucma_finish_ctx(ctx);
 319	list_add_tail(&uevent->list, &ctx->file->event_list);
 320	mutex_unlock(&ctx->file->mut);
 321	wake_up_interruptible(&ctx->file->poll_wait);
 322	return 0;
 
 
 
 
 
 
 
 323
 324err_alloc:
 325	ucma_destroy_private_ctx(ctx);
 326err_backlog:
 327	atomic_inc(&listen_ctx->backlog);
 328	/* Returning error causes the new ID to be destroyed */
 329	return -ENOMEM;
 
 
 
 
 
 
 330}
 331
 332static int ucma_event_handler(struct rdma_cm_id *cm_id,
 333			      struct rdma_cm_event *event)
 334{
 335	struct ucma_event *uevent;
 336	struct ucma_context *ctx = cm_id->context;
 
 337
 338	if (event->event == RDMA_CM_EVENT_CONNECT_REQUEST)
 339		return ucma_connect_event_handler(cm_id, event);
 
 340
 341	/*
 342	 * We ignore events for new connections until userspace has set their
 343	 * context.  This can only happen if an error occurs on a new connection
 344	 * before the user accepts it.  This is okay, since the accept will just
 345	 * fail later. However, we do need to release the underlying HW
 346	 * resources in case of a device removal event.
 347	 */
 348	if (ctx->uid) {
 349		uevent = ucma_create_uevent(ctx, event);
 350		if (!uevent)
 351			return 0;
 352
 353		mutex_lock(&ctx->file->mut);
 354		list_add_tail(&uevent->list, &ctx->file->event_list);
 355		mutex_unlock(&ctx->file->mut);
 356		wake_up_interruptible(&ctx->file->poll_wait);
 357	}
 358
 359	if (event->event == RDMA_CM_EVENT_DEVICE_REMOVAL) {
 360		xa_lock(&ctx_table);
 361		if (xa_load(&ctx_table, ctx->id) == ctx)
 362			queue_work(system_unbound_wq, &ctx->close_work);
 363		xa_unlock(&ctx_table);
 
 
 
 
 
 
 
 
 
 
 364	}
 365	return 0;
 
 
 
 
 
 
 
 366}
 367
 368static ssize_t ucma_get_event(struct ucma_file *file, const char __user *inbuf,
 369			      int in_len, int out_len)
 370{
 
 371	struct rdma_ucm_get_event cmd;
 372	struct ucma_event *uevent;
 
 373
 374	/*
 375	 * Old 32 bit user space does not send the 4 byte padding in the
 376	 * reserved field. We don't care, allow it to keep working.
 377	 */
 378	if (out_len < sizeof(uevent->resp) - sizeof(uevent->resp.reserved) -
 379			      sizeof(uevent->resp.ece))
 380		return -ENOSPC;
 381
 382	if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
 383		return -EFAULT;
 384
 385	mutex_lock(&file->mut);
 386	while (list_empty(&file->event_list)) {
 387		mutex_unlock(&file->mut);
 388
 389		if (file->filp->f_flags & O_NONBLOCK)
 390			return -EAGAIN;
 391
 392		if (wait_event_interruptible(file->poll_wait,
 393					     !list_empty(&file->event_list)))
 394			return -ERESTARTSYS;
 395
 396		mutex_lock(&file->mut);
 397	}
 398
 399	uevent = list_first_entry(&file->event_list, struct ucma_event, list);
 
 
 
 
 
 
 
 
 
 
 
 
 400
 401	if (copy_to_user(u64_to_user_ptr(cmd.response),
 402			 &uevent->resp,
 403			 min_t(size_t, out_len, sizeof(uevent->resp)))) {
 404		mutex_unlock(&file->mut);
 405		return -EFAULT;
 406	}
 407
 408	list_del(&uevent->list);
 409	uevent->ctx->events_reported++;
 410	if (uevent->mc)
 411		uevent->mc->events_reported++;
 412	if (uevent->resp.event == RDMA_CM_EVENT_CONNECT_REQUEST)
 413		atomic_inc(&uevent->ctx->backlog);
 414	mutex_unlock(&file->mut);
 415
 416	kfree(uevent);
 417	return 0;
 
 
 418}
 419
 420static int ucma_get_qp_type(struct rdma_ucm_create_id *cmd, enum ib_qp_type *qp_type)
 421{
 422	switch (cmd->ps) {
 423	case RDMA_PS_TCP:
 424		*qp_type = IB_QPT_RC;
 425		return 0;
 426	case RDMA_PS_UDP:
 427	case RDMA_PS_IPOIB:
 428		*qp_type = IB_QPT_UD;
 429		return 0;
 430	case RDMA_PS_IB:
 431		*qp_type = cmd->qp_type;
 432		return 0;
 433	default:
 434		return -EINVAL;
 435	}
 436}
 437
 438static ssize_t ucma_create_id(struct ucma_file *file, const char __user *inbuf,
 439			      int in_len, int out_len)
 440{
 441	struct rdma_ucm_create_id cmd;
 442	struct rdma_ucm_create_id_resp resp;
 443	struct ucma_context *ctx;
 444	struct rdma_cm_id *cm_id;
 445	enum ib_qp_type qp_type;
 446	int ret;
 447
 448	if (out_len < sizeof(resp))
 449		return -ENOSPC;
 450
 451	if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
 452		return -EFAULT;
 453
 454	ret = ucma_get_qp_type(&cmd, &qp_type);
 455	if (ret)
 456		return ret;
 457
 
 458	ctx = ucma_alloc_ctx(file);
 
 459	if (!ctx)
 460		return -ENOMEM;
 461
 462	ctx->uid = cmd.uid;
 463	cm_id = rdma_create_user_id(ucma_event_handler, ctx, cmd.ps, qp_type);
 
 464	if (IS_ERR(cm_id)) {
 465		ret = PTR_ERR(cm_id);
 466		goto err1;
 467	}
 468	ucma_set_ctx_cm_id(ctx, cm_id);
 469
 470	resp.id = ctx->id;
 471	if (copy_to_user(u64_to_user_ptr(cmd.response),
 472			 &resp, sizeof(resp))) {
 473		ret = -EFAULT;
 474		goto err1;
 475	}
 476
 477	mutex_lock(&file->mut);
 478	ucma_finish_ctx(ctx);
 479	mutex_unlock(&file->mut);
 480	return 0;
 481
 
 
 482err1:
 483	ucma_destroy_private_ctx(ctx);
 
 
 
 
 
 
 484	return ret;
 485}
 486
 487static void ucma_cleanup_multicast(struct ucma_context *ctx)
 488{
 489	struct ucma_multicast *mc, *tmp;
 490
 491	xa_lock(&multicast_table);
 492	list_for_each_entry_safe(mc, tmp, &ctx->mc_list, list) {
 493		list_del(&mc->list);
 494		/*
 495		 * At this point mc->ctx->ref is 0 so the mc cannot leave the
 496		 * lock on the reader and this is enough serialization
 497		 */
 498		__xa_erase(&multicast_table, mc->id);
 499		kfree(mc);
 500	}
 501	xa_unlock(&multicast_table);
 502}
 503
 504static void ucma_cleanup_mc_events(struct ucma_multicast *mc)
 505{
 506	struct ucma_event *uevent, *tmp;
 507
 508	rdma_lock_handler(mc->ctx->cm_id);
 509	mutex_lock(&mc->ctx->file->mut);
 510	list_for_each_entry_safe(uevent, tmp, &mc->ctx->file->event_list, list) {
 511		if (uevent->mc != mc)
 512			continue;
 513
 514		list_del(&uevent->list);
 515		kfree(uevent);
 516	}
 517	mutex_unlock(&mc->ctx->file->mut);
 518	rdma_unlock_handler(mc->ctx->cm_id);
 519}
 520
 521static int ucma_cleanup_ctx_events(struct ucma_context *ctx)
 
 
 
 
 
 
 
 
 
 
 
 522{
 523	int events_reported;
 524	struct ucma_event *uevent, *tmp;
 525	LIST_HEAD(list);
 526
 527	/* Cleanup events not yet reported to the user.*/
 
 
 
 528	mutex_lock(&ctx->file->mut);
 529	list_for_each_entry_safe(uevent, tmp, &ctx->file->event_list, list) {
 530		if (uevent->ctx != ctx)
 531			continue;
 532
 533		if (uevent->resp.event == RDMA_CM_EVENT_CONNECT_REQUEST &&
 534		    xa_cmpxchg(&ctx_table, uevent->conn_req_ctx->id,
 535			       uevent->conn_req_ctx, XA_ZERO_ENTRY,
 536			       GFP_KERNEL) == uevent->conn_req_ctx) {
 537			list_move_tail(&uevent->list, &list);
 538			continue;
 539		}
 540		list_del(&uevent->list);
 541		kfree(uevent);
 542	}
 543	list_del(&ctx->list);
 544	events_reported = ctx->events_reported;
 545	mutex_unlock(&ctx->file->mut);
 546
 547	/*
 548	 * If this was a listening ID then any connections spawned from it that
 549	 * have not been delivered to userspace are cleaned up too. Must be done
 550	 * outside any locks.
 551	 */
 552	list_for_each_entry_safe(uevent, tmp, &list, list) {
 553		ucma_destroy_private_ctx(uevent->conn_req_ctx);
 
 
 554		kfree(uevent);
 555	}
 556	return events_reported;
 557}
 558
 559/*
 560 * When this is called the xarray must have a XA_ZERO_ENTRY in the ctx->id (ie
 561 * the ctx is not public to the user). This either because:
 562 *  - ucma_finish_ctx() hasn't been called
 563 *  - xa_cmpxchg() succeed to remove the entry (only one thread can succeed)
 564 */
 565static int ucma_destroy_private_ctx(struct ucma_context *ctx)
 566{
 567	int events_reported;
 568
 569	/*
 570	 * Destroy the underlying cm_id. New work queuing is prevented now by
 571	 * the removal from the xarray. Once the work is cancled ref will either
 572	 * be 0 because the work ran to completion and consumed the ref from the
 573	 * xarray, or it will be positive because we still have the ref from the
 574	 * xarray. This can also be 0 in cases where cm_id was never set
 575	 */
 576	cancel_work_sync(&ctx->close_work);
 577	if (refcount_read(&ctx->ref))
 578		ucma_close_id(&ctx->close_work);
 579
 580	events_reported = ucma_cleanup_ctx_events(ctx);
 581	ucma_cleanup_multicast(ctx);
 582
 583	WARN_ON(xa_cmpxchg(&ctx_table, ctx->id, XA_ZERO_ENTRY, NULL,
 584			   GFP_KERNEL) != NULL);
 585	mutex_destroy(&ctx->mutex);
 586	kfree(ctx);
 587	return events_reported;
 588}
 589
 590static ssize_t ucma_destroy_id(struct ucma_file *file, const char __user *inbuf,
 591			       int in_len, int out_len)
 592{
 593	struct rdma_ucm_destroy_id cmd;
 594	struct rdma_ucm_destroy_id_resp resp;
 595	struct ucma_context *ctx;
 596	int ret = 0;
 597
 598	if (out_len < sizeof(resp))
 599		return -ENOSPC;
 600
 601	if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
 602		return -EFAULT;
 603
 604	xa_lock(&ctx_table);
 605	ctx = _ucma_find_context(cmd.id, file);
 606	if (!IS_ERR(ctx)) {
 607		if (__xa_cmpxchg(&ctx_table, ctx->id, ctx, XA_ZERO_ENTRY,
 608				 GFP_KERNEL) != ctx)
 609			ctx = ERR_PTR(-ENOENT);
 610	}
 611	xa_unlock(&ctx_table);
 612
 613	if (IS_ERR(ctx))
 614		return PTR_ERR(ctx);
 615
 616	resp.events_reported = ucma_destroy_private_ctx(ctx);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 617	if (copy_to_user(u64_to_user_ptr(cmd.response),
 618			 &resp, sizeof(resp)))
 619		ret = -EFAULT;
 620
 621	return ret;
 622}
 623
 624static ssize_t ucma_bind_ip(struct ucma_file *file, const char __user *inbuf,
 625			      int in_len, int out_len)
 626{
 627	struct rdma_ucm_bind_ip cmd;
 628	struct ucma_context *ctx;
 629	int ret;
 630
 631	if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
 632		return -EFAULT;
 633
 634	if (!rdma_addr_size_in6(&cmd.addr))
 635		return -EINVAL;
 636
 637	ctx = ucma_get_ctx(file, cmd.id);
 638	if (IS_ERR(ctx))
 639		return PTR_ERR(ctx);
 640
 641	mutex_lock(&ctx->mutex);
 642	ret = rdma_bind_addr(ctx->cm_id, (struct sockaddr *) &cmd.addr);
 643	mutex_unlock(&ctx->mutex);
 644
 645	ucma_put_ctx(ctx);
 646	return ret;
 647}
 648
 649static ssize_t ucma_bind(struct ucma_file *file, const char __user *inbuf,
 650			 int in_len, int out_len)
 651{
 652	struct rdma_ucm_bind cmd;
 653	struct ucma_context *ctx;
 654	int ret;
 655
 656	if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
 657		return -EFAULT;
 658
 659	if (cmd.reserved || !cmd.addr_size ||
 660	    cmd.addr_size != rdma_addr_size_kss(&cmd.addr))
 661		return -EINVAL;
 662
 663	ctx = ucma_get_ctx(file, cmd.id);
 664	if (IS_ERR(ctx))
 665		return PTR_ERR(ctx);
 666
 667	mutex_lock(&ctx->mutex);
 668	ret = rdma_bind_addr(ctx->cm_id, (struct sockaddr *) &cmd.addr);
 669	mutex_unlock(&ctx->mutex);
 670	ucma_put_ctx(ctx);
 671	return ret;
 672}
 673
 674static ssize_t ucma_resolve_ip(struct ucma_file *file,
 675			       const char __user *inbuf,
 676			       int in_len, int out_len)
 677{
 678	struct rdma_ucm_resolve_ip cmd;
 679	struct ucma_context *ctx;
 680	int ret;
 681
 682	if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
 683		return -EFAULT;
 684
 685	if ((cmd.src_addr.sin6_family && !rdma_addr_size_in6(&cmd.src_addr)) ||
 686	    !rdma_addr_size_in6(&cmd.dst_addr))
 687		return -EINVAL;
 688
 689	ctx = ucma_get_ctx(file, cmd.id);
 690	if (IS_ERR(ctx))
 691		return PTR_ERR(ctx);
 692
 693	mutex_lock(&ctx->mutex);
 694	ret = rdma_resolve_addr(ctx->cm_id, (struct sockaddr *) &cmd.src_addr,
 695				(struct sockaddr *) &cmd.dst_addr, cmd.timeout_ms);
 696	mutex_unlock(&ctx->mutex);
 697	ucma_put_ctx(ctx);
 698	return ret;
 699}
 700
 701static ssize_t ucma_resolve_addr(struct ucma_file *file,
 702				 const char __user *inbuf,
 703				 int in_len, int out_len)
 704{
 705	struct rdma_ucm_resolve_addr cmd;
 706	struct ucma_context *ctx;
 707	int ret;
 708
 709	if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
 710		return -EFAULT;
 711
 712	if (cmd.reserved ||
 713	    (cmd.src_size && (cmd.src_size != rdma_addr_size_kss(&cmd.src_addr))) ||
 714	    !cmd.dst_size || (cmd.dst_size != rdma_addr_size_kss(&cmd.dst_addr)))
 715		return -EINVAL;
 716
 717	ctx = ucma_get_ctx(file, cmd.id);
 718	if (IS_ERR(ctx))
 719		return PTR_ERR(ctx);
 720
 721	mutex_lock(&ctx->mutex);
 722	ret = rdma_resolve_addr(ctx->cm_id, (struct sockaddr *) &cmd.src_addr,
 723				(struct sockaddr *) &cmd.dst_addr, cmd.timeout_ms);
 724	mutex_unlock(&ctx->mutex);
 725	ucma_put_ctx(ctx);
 726	return ret;
 727}
 728
 729static ssize_t ucma_resolve_route(struct ucma_file *file,
 730				  const char __user *inbuf,
 731				  int in_len, int out_len)
 732{
 733	struct rdma_ucm_resolve_route cmd;
 734	struct ucma_context *ctx;
 735	int ret;
 736
 737	if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
 738		return -EFAULT;
 739
 740	ctx = ucma_get_ctx_dev(file, cmd.id);
 741	if (IS_ERR(ctx))
 742		return PTR_ERR(ctx);
 743
 744	mutex_lock(&ctx->mutex);
 745	ret = rdma_resolve_route(ctx->cm_id, cmd.timeout_ms);
 746	mutex_unlock(&ctx->mutex);
 747	ucma_put_ctx(ctx);
 748	return ret;
 749}
 750
 751static void ucma_copy_ib_route(struct rdma_ucm_query_route_resp *resp,
 752			       struct rdma_route *route)
 753{
 754	struct rdma_dev_addr *dev_addr;
 755
 756	resp->num_paths = route->num_pri_alt_paths;
 757	switch (route->num_pri_alt_paths) {
 758	case 0:
 759		dev_addr = &route->addr.dev_addr;
 760		rdma_addr_get_dgid(dev_addr,
 761				   (union ib_gid *) &resp->ib_route[0].dgid);
 762		rdma_addr_get_sgid(dev_addr,
 763				   (union ib_gid *) &resp->ib_route[0].sgid);
 764		resp->ib_route[0].pkey = cpu_to_be16(ib_addr_get_pkey(dev_addr));
 765		break;
 766	case 2:
 767		ib_copy_path_rec_to_user(&resp->ib_route[1],
 768					 &route->path_rec[1]);
 769		fallthrough;
 770	case 1:
 771		ib_copy_path_rec_to_user(&resp->ib_route[0],
 772					 &route->path_rec[0]);
 773		break;
 774	default:
 775		break;
 776	}
 777}
 778
 779static void ucma_copy_iboe_route(struct rdma_ucm_query_route_resp *resp,
 780				 struct rdma_route *route)
 781{
 782
 783	resp->num_paths = route->num_pri_alt_paths;
 784	switch (route->num_pri_alt_paths) {
 785	case 0:
 786		rdma_ip2gid((struct sockaddr *)&route->addr.dst_addr,
 787			    (union ib_gid *)&resp->ib_route[0].dgid);
 788		rdma_ip2gid((struct sockaddr *)&route->addr.src_addr,
 789			    (union ib_gid *)&resp->ib_route[0].sgid);
 790		resp->ib_route[0].pkey = cpu_to_be16(0xffff);
 791		break;
 792	case 2:
 793		ib_copy_path_rec_to_user(&resp->ib_route[1],
 794					 &route->path_rec[1]);
 795		fallthrough;
 796	case 1:
 797		ib_copy_path_rec_to_user(&resp->ib_route[0],
 798					 &route->path_rec[0]);
 799		break;
 800	default:
 801		break;
 802	}
 803}
 804
 805static void ucma_copy_iw_route(struct rdma_ucm_query_route_resp *resp,
 806			       struct rdma_route *route)
 807{
 808	struct rdma_dev_addr *dev_addr;
 809
 810	dev_addr = &route->addr.dev_addr;
 811	rdma_addr_get_dgid(dev_addr, (union ib_gid *) &resp->ib_route[0].dgid);
 812	rdma_addr_get_sgid(dev_addr, (union ib_gid *) &resp->ib_route[0].sgid);
 813}
 814
 815static ssize_t ucma_query_route(struct ucma_file *file,
 816				const char __user *inbuf,
 817				int in_len, int out_len)
 818{
 819	struct rdma_ucm_query cmd;
 820	struct rdma_ucm_query_route_resp resp;
 821	struct ucma_context *ctx;
 822	struct sockaddr *addr;
 823	int ret = 0;
 824
 825	if (out_len < offsetof(struct rdma_ucm_query_route_resp, ibdev_index))
 826		return -ENOSPC;
 827
 828	if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
 829		return -EFAULT;
 830
 831	ctx = ucma_get_ctx(file, cmd.id);
 832	if (IS_ERR(ctx))
 833		return PTR_ERR(ctx);
 834
 835	mutex_lock(&ctx->mutex);
 836	memset(&resp, 0, sizeof resp);
 837	addr = (struct sockaddr *) &ctx->cm_id->route.addr.src_addr;
 838	memcpy(&resp.src_addr, addr, addr->sa_family == AF_INET ?
 839				     sizeof(struct sockaddr_in) :
 840				     sizeof(struct sockaddr_in6));
 841	addr = (struct sockaddr *) &ctx->cm_id->route.addr.dst_addr;
 842	memcpy(&resp.dst_addr, addr, addr->sa_family == AF_INET ?
 843				     sizeof(struct sockaddr_in) :
 844				     sizeof(struct sockaddr_in6));
 845	if (!ctx->cm_id->device)
 846		goto out;
 847
 848	resp.node_guid = (__force __u64) ctx->cm_id->device->node_guid;
 849	resp.ibdev_index = ctx->cm_id->device->index;
 850	resp.port_num = ctx->cm_id->port_num;
 851
 852	if (rdma_cap_ib_sa(ctx->cm_id->device, ctx->cm_id->port_num))
 853		ucma_copy_ib_route(&resp, &ctx->cm_id->route);
 854	else if (rdma_protocol_roce(ctx->cm_id->device, ctx->cm_id->port_num))
 855		ucma_copy_iboe_route(&resp, &ctx->cm_id->route);
 856	else if (rdma_protocol_iwarp(ctx->cm_id->device, ctx->cm_id->port_num))
 857		ucma_copy_iw_route(&resp, &ctx->cm_id->route);
 858
 859out:
 860	mutex_unlock(&ctx->mutex);
 861	if (copy_to_user(u64_to_user_ptr(cmd.response), &resp,
 862			 min_t(size_t, out_len, sizeof(resp))))
 863		ret = -EFAULT;
 864
 865	ucma_put_ctx(ctx);
 866	return ret;
 867}
 868
 869static void ucma_query_device_addr(struct rdma_cm_id *cm_id,
 870				   struct rdma_ucm_query_addr_resp *resp)
 871{
 872	if (!cm_id->device)
 873		return;
 874
 875	resp->node_guid = (__force __u64) cm_id->device->node_guid;
 876	resp->ibdev_index = cm_id->device->index;
 877	resp->port_num = cm_id->port_num;
 878	resp->pkey = (__force __u16) cpu_to_be16(
 879		     ib_addr_get_pkey(&cm_id->route.addr.dev_addr));
 880}
 881
 882static ssize_t ucma_query_addr(struct ucma_context *ctx,
 883			       void __user *response, int out_len)
 884{
 885	struct rdma_ucm_query_addr_resp resp;
 886	struct sockaddr *addr;
 887	int ret = 0;
 888
 889	if (out_len < offsetof(struct rdma_ucm_query_addr_resp, ibdev_index))
 890		return -ENOSPC;
 891
 892	memset(&resp, 0, sizeof resp);
 893
 894	addr = (struct sockaddr *) &ctx->cm_id->route.addr.src_addr;
 895	resp.src_size = rdma_addr_size(addr);
 896	memcpy(&resp.src_addr, addr, resp.src_size);
 897
 898	addr = (struct sockaddr *) &ctx->cm_id->route.addr.dst_addr;
 899	resp.dst_size = rdma_addr_size(addr);
 900	memcpy(&resp.dst_addr, addr, resp.dst_size);
 901
 902	ucma_query_device_addr(ctx->cm_id, &resp);
 903
 904	if (copy_to_user(response, &resp, min_t(size_t, out_len, sizeof(resp))))
 905		ret = -EFAULT;
 906
 907	return ret;
 908}
 909
 910static ssize_t ucma_query_path(struct ucma_context *ctx,
 911			       void __user *response, int out_len)
 912{
 913	struct rdma_ucm_query_path_resp *resp;
 914	int i, ret = 0;
 915
 916	if (out_len < sizeof(*resp))
 917		return -ENOSPC;
 918
 919	resp = kzalloc(out_len, GFP_KERNEL);
 920	if (!resp)
 921		return -ENOMEM;
 922
 923	resp->num_paths = ctx->cm_id->route.num_pri_alt_paths;
 924	for (i = 0, out_len -= sizeof(*resp);
 925	     i < resp->num_paths && out_len > sizeof(struct ib_path_rec_data);
 926	     i++, out_len -= sizeof(struct ib_path_rec_data)) {
 927		struct sa_path_rec *rec = &ctx->cm_id->route.path_rec[i];
 928
 929		resp->path_data[i].flags = IB_PATH_GMP | IB_PATH_PRIMARY |
 930					   IB_PATH_BIDIRECTIONAL;
 931		if (rec->rec_type == SA_PATH_REC_TYPE_OPA) {
 932			struct sa_path_rec ib;
 933
 934			sa_convert_path_opa_to_ib(&ib, rec);
 935			ib_sa_pack_path(&ib, &resp->path_data[i].path_rec);
 936
 937		} else {
 938			ib_sa_pack_path(rec, &resp->path_data[i].path_rec);
 939		}
 940	}
 941
 942	if (copy_to_user(response, resp, struct_size(resp, path_data, i)))
 
 943		ret = -EFAULT;
 944
 945	kfree(resp);
 946	return ret;
 947}
 948
 949static ssize_t ucma_query_gid(struct ucma_context *ctx,
 950			      void __user *response, int out_len)
 951{
 952	struct rdma_ucm_query_addr_resp resp;
 953	struct sockaddr_ib *addr;
 954	int ret = 0;
 955
 956	if (out_len < offsetof(struct rdma_ucm_query_addr_resp, ibdev_index))
 957		return -ENOSPC;
 958
 959	memset(&resp, 0, sizeof resp);
 960
 961	ucma_query_device_addr(ctx->cm_id, &resp);
 962
 963	addr = (struct sockaddr_ib *) &resp.src_addr;
 964	resp.src_size = sizeof(*addr);
 965	if (ctx->cm_id->route.addr.src_addr.ss_family == AF_IB) {
 966		memcpy(addr, &ctx->cm_id->route.addr.src_addr, resp.src_size);
 967	} else {
 968		addr->sib_family = AF_IB;
 969		addr->sib_pkey = (__force __be16) resp.pkey;
 970		rdma_read_gids(ctx->cm_id, (union ib_gid *)&addr->sib_addr,
 971			       NULL);
 972		addr->sib_sid = rdma_get_service_id(ctx->cm_id, (struct sockaddr *)
 973						    &ctx->cm_id->route.addr.src_addr);
 974	}
 975
 976	addr = (struct sockaddr_ib *) &resp.dst_addr;
 977	resp.dst_size = sizeof(*addr);
 978	if (ctx->cm_id->route.addr.dst_addr.ss_family == AF_IB) {
 979		memcpy(addr, &ctx->cm_id->route.addr.dst_addr, resp.dst_size);
 980	} else {
 981		addr->sib_family = AF_IB;
 982		addr->sib_pkey = (__force __be16) resp.pkey;
 983		rdma_read_gids(ctx->cm_id, NULL,
 984			       (union ib_gid *)&addr->sib_addr);
 985		addr->sib_sid = rdma_get_service_id(ctx->cm_id, (struct sockaddr *)
 986						    &ctx->cm_id->route.addr.dst_addr);
 987	}
 988
 989	if (copy_to_user(response, &resp, min_t(size_t, out_len, sizeof(resp))))
 990		ret = -EFAULT;
 991
 992	return ret;
 993}
 994
 995static ssize_t ucma_query(struct ucma_file *file,
 996			  const char __user *inbuf,
 997			  int in_len, int out_len)
 998{
 999	struct rdma_ucm_query cmd;
1000	struct ucma_context *ctx;
1001	void __user *response;
1002	int ret;
1003
1004	if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
1005		return -EFAULT;
1006
1007	response = u64_to_user_ptr(cmd.response);
1008	ctx = ucma_get_ctx(file, cmd.id);
1009	if (IS_ERR(ctx))
1010		return PTR_ERR(ctx);
1011
1012	mutex_lock(&ctx->mutex);
1013	switch (cmd.option) {
1014	case RDMA_USER_CM_QUERY_ADDR:
1015		ret = ucma_query_addr(ctx, response, out_len);
1016		break;
1017	case RDMA_USER_CM_QUERY_PATH:
1018		ret = ucma_query_path(ctx, response, out_len);
1019		break;
1020	case RDMA_USER_CM_QUERY_GID:
1021		ret = ucma_query_gid(ctx, response, out_len);
1022		break;
1023	default:
1024		ret = -ENOSYS;
1025		break;
1026	}
1027	mutex_unlock(&ctx->mutex);
1028
1029	ucma_put_ctx(ctx);
1030	return ret;
1031}
1032
1033static void ucma_copy_conn_param(struct rdma_cm_id *id,
1034				 struct rdma_conn_param *dst,
1035				 struct rdma_ucm_conn_param *src)
1036{
1037	dst->private_data = src->private_data;
1038	dst->private_data_len = src->private_data_len;
1039	dst->responder_resources = src->responder_resources;
1040	dst->initiator_depth = src->initiator_depth;
1041	dst->flow_control = src->flow_control;
1042	dst->retry_count = src->retry_count;
1043	dst->rnr_retry_count = src->rnr_retry_count;
1044	dst->srq = src->srq;
1045	dst->qp_num = src->qp_num & 0xFFFFFF;
1046	dst->qkey = (id->route.addr.src_addr.ss_family == AF_IB) ? src->qkey : 0;
1047}
1048
1049static ssize_t ucma_connect(struct ucma_file *file, const char __user *inbuf,
1050			    int in_len, int out_len)
1051{
1052	struct rdma_conn_param conn_param;
1053	struct rdma_ucm_ece ece = {};
1054	struct rdma_ucm_connect cmd;
 
1055	struct ucma_context *ctx;
1056	size_t in_size;
1057	int ret;
1058
1059	if (in_len < offsetofend(typeof(cmd), reserved))
1060		return -EINVAL;
1061	in_size = min_t(size_t, in_len, sizeof(cmd));
1062	if (copy_from_user(&cmd, inbuf, in_size))
1063		return -EFAULT;
1064
1065	if (!cmd.conn_param.valid)
1066		return -EINVAL;
1067
1068	ctx = ucma_get_ctx_dev(file, cmd.id);
1069	if (IS_ERR(ctx))
1070		return PTR_ERR(ctx);
1071
1072	ucma_copy_conn_param(ctx->cm_id, &conn_param, &cmd.conn_param);
1073	if (offsetofend(typeof(cmd), ece) <= in_size) {
1074		ece.vendor_id = cmd.ece.vendor_id;
1075		ece.attr_mod = cmd.ece.attr_mod;
1076	}
1077
1078	mutex_lock(&ctx->mutex);
1079	ret = rdma_connect_ece(ctx->cm_id, &conn_param, &ece);
1080	mutex_unlock(&ctx->mutex);
1081	ucma_put_ctx(ctx);
1082	return ret;
1083}
1084
1085static ssize_t ucma_listen(struct ucma_file *file, const char __user *inbuf,
1086			   int in_len, int out_len)
1087{
1088	struct rdma_ucm_listen cmd;
1089	struct ucma_context *ctx;
1090	int ret;
1091
1092	if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
1093		return -EFAULT;
1094
1095	ctx = ucma_get_ctx(file, cmd.id);
1096	if (IS_ERR(ctx))
1097		return PTR_ERR(ctx);
1098
1099	if (cmd.backlog <= 0 || cmd.backlog > max_backlog)
1100		cmd.backlog = max_backlog;
1101	atomic_set(&ctx->backlog, cmd.backlog);
1102
1103	mutex_lock(&ctx->mutex);
1104	ret = rdma_listen(ctx->cm_id, cmd.backlog);
1105	mutex_unlock(&ctx->mutex);
1106	ucma_put_ctx(ctx);
1107	return ret;
1108}
1109
1110static ssize_t ucma_accept(struct ucma_file *file, const char __user *inbuf,
1111			   int in_len, int out_len)
1112{
1113	struct rdma_ucm_accept cmd;
1114	struct rdma_conn_param conn_param;
1115	struct rdma_ucm_ece ece = {};
1116	struct ucma_context *ctx;
1117	size_t in_size;
1118	int ret;
1119
1120	if (in_len < offsetofend(typeof(cmd), reserved))
1121		return -EINVAL;
1122	in_size = min_t(size_t, in_len, sizeof(cmd));
1123	if (copy_from_user(&cmd, inbuf, in_size))
1124		return -EFAULT;
1125
1126	ctx = ucma_get_ctx_dev(file, cmd.id);
1127	if (IS_ERR(ctx))
1128		return PTR_ERR(ctx);
1129
1130	if (offsetofend(typeof(cmd), ece) <= in_size) {
1131		ece.vendor_id = cmd.ece.vendor_id;
1132		ece.attr_mod = cmd.ece.attr_mod;
1133	}
1134
1135	if (cmd.conn_param.valid) {
1136		ucma_copy_conn_param(ctx->cm_id, &conn_param, &cmd.conn_param);
1137		mutex_lock(&ctx->mutex);
1138		rdma_lock_handler(ctx->cm_id);
1139		ret = rdma_accept_ece(ctx->cm_id, &conn_param, &ece);
1140		if (!ret) {
1141			/* The uid must be set atomically with the handler */
1142			ctx->uid = cmd.uid;
1143		}
1144		rdma_unlock_handler(ctx->cm_id);
1145		mutex_unlock(&ctx->mutex);
1146	} else {
1147		mutex_lock(&ctx->mutex);
1148		rdma_lock_handler(ctx->cm_id);
1149		ret = rdma_accept_ece(ctx->cm_id, NULL, &ece);
1150		rdma_unlock_handler(ctx->cm_id);
1151		mutex_unlock(&ctx->mutex);
1152	}
1153	ucma_put_ctx(ctx);
1154	return ret;
1155}
1156
1157static ssize_t ucma_reject(struct ucma_file *file, const char __user *inbuf,
1158			   int in_len, int out_len)
1159{
1160	struct rdma_ucm_reject cmd;
1161	struct ucma_context *ctx;
1162	int ret;
1163
1164	if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
1165		return -EFAULT;
1166
1167	if (!cmd.reason)
1168		cmd.reason = IB_CM_REJ_CONSUMER_DEFINED;
1169
1170	switch (cmd.reason) {
1171	case IB_CM_REJ_CONSUMER_DEFINED:
1172	case IB_CM_REJ_VENDOR_OPTION_NOT_SUPPORTED:
1173		break;
1174	default:
1175		return -EINVAL;
1176	}
1177
1178	ctx = ucma_get_ctx_dev(file, cmd.id);
1179	if (IS_ERR(ctx))
1180		return PTR_ERR(ctx);
1181
1182	mutex_lock(&ctx->mutex);
1183	ret = rdma_reject(ctx->cm_id, cmd.private_data, cmd.private_data_len,
1184			  cmd.reason);
1185	mutex_unlock(&ctx->mutex);
1186	ucma_put_ctx(ctx);
1187	return ret;
1188}
1189
1190static ssize_t ucma_disconnect(struct ucma_file *file, const char __user *inbuf,
1191			       int in_len, int out_len)
1192{
1193	struct rdma_ucm_disconnect cmd;
1194	struct ucma_context *ctx;
1195	int ret;
1196
1197	if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
1198		return -EFAULT;
1199
1200	ctx = ucma_get_ctx_dev(file, cmd.id);
1201	if (IS_ERR(ctx))
1202		return PTR_ERR(ctx);
1203
1204	mutex_lock(&ctx->mutex);
1205	ret = rdma_disconnect(ctx->cm_id);
1206	mutex_unlock(&ctx->mutex);
1207	ucma_put_ctx(ctx);
1208	return ret;
1209}
1210
1211static ssize_t ucma_init_qp_attr(struct ucma_file *file,
1212				 const char __user *inbuf,
1213				 int in_len, int out_len)
1214{
1215	struct rdma_ucm_init_qp_attr cmd;
1216	struct ib_uverbs_qp_attr resp;
1217	struct ucma_context *ctx;
1218	struct ib_qp_attr qp_attr;
1219	int ret;
1220
1221	if (out_len < sizeof(resp))
1222		return -ENOSPC;
1223
1224	if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
1225		return -EFAULT;
1226
1227	if (cmd.qp_state > IB_QPS_ERR)
1228		return -EINVAL;
1229
1230	ctx = ucma_get_ctx_dev(file, cmd.id);
1231	if (IS_ERR(ctx))
1232		return PTR_ERR(ctx);
1233
1234	resp.qp_attr_mask = 0;
1235	memset(&qp_attr, 0, sizeof qp_attr);
1236	qp_attr.qp_state = cmd.qp_state;
1237	mutex_lock(&ctx->mutex);
1238	ret = rdma_init_qp_attr(ctx->cm_id, &qp_attr, &resp.qp_attr_mask);
1239	mutex_unlock(&ctx->mutex);
1240	if (ret)
1241		goto out;
1242
1243	ib_copy_qp_attr_to_user(ctx->cm_id->device, &resp, &qp_attr);
1244	if (copy_to_user(u64_to_user_ptr(cmd.response),
1245			 &resp, sizeof(resp)))
1246		ret = -EFAULT;
1247
1248out:
1249	ucma_put_ctx(ctx);
1250	return ret;
1251}
1252
1253static int ucma_set_option_id(struct ucma_context *ctx, int optname,
1254			      void *optval, size_t optlen)
1255{
1256	int ret = 0;
1257
1258	switch (optname) {
1259	case RDMA_OPTION_ID_TOS:
1260		if (optlen != sizeof(u8)) {
1261			ret = -EINVAL;
1262			break;
1263		}
1264		rdma_set_service_type(ctx->cm_id, *((u8 *) optval));
1265		break;
1266	case RDMA_OPTION_ID_REUSEADDR:
1267		if (optlen != sizeof(int)) {
1268			ret = -EINVAL;
1269			break;
1270		}
1271		ret = rdma_set_reuseaddr(ctx->cm_id, *((int *) optval) ? 1 : 0);
1272		break;
1273	case RDMA_OPTION_ID_AFONLY:
1274		if (optlen != sizeof(int)) {
1275			ret = -EINVAL;
1276			break;
1277		}
1278		ret = rdma_set_afonly(ctx->cm_id, *((int *) optval) ? 1 : 0);
1279		break;
1280	case RDMA_OPTION_ID_ACK_TIMEOUT:
1281		if (optlen != sizeof(u8)) {
1282			ret = -EINVAL;
1283			break;
1284		}
1285		ret = rdma_set_ack_timeout(ctx->cm_id, *((u8 *)optval));
1286		break;
1287	default:
1288		ret = -ENOSYS;
1289	}
1290
1291	return ret;
1292}
1293
1294static int ucma_set_ib_path(struct ucma_context *ctx,
1295			    struct ib_path_rec_data *path_data, size_t optlen)
1296{
1297	struct sa_path_rec sa_path;
1298	struct rdma_cm_event event;
1299	int ret;
1300
1301	if (optlen % sizeof(*path_data))
1302		return -EINVAL;
1303
1304	for (; optlen; optlen -= sizeof(*path_data), path_data++) {
1305		if (path_data->flags == (IB_PATH_GMP | IB_PATH_PRIMARY |
1306					 IB_PATH_BIDIRECTIONAL))
1307			break;
1308	}
1309
1310	if (!optlen)
1311		return -EINVAL;
1312
1313	if (!ctx->cm_id->device)
1314		return -EINVAL;
1315
1316	memset(&sa_path, 0, sizeof(sa_path));
1317
1318	sa_path.rec_type = SA_PATH_REC_TYPE_IB;
1319	ib_sa_unpack_path(path_data->path_rec, &sa_path);
1320
1321	if (rdma_cap_opa_ah(ctx->cm_id->device, ctx->cm_id->port_num)) {
1322		struct sa_path_rec opa;
1323
1324		sa_convert_path_ib_to_opa(&opa, &sa_path);
1325		mutex_lock(&ctx->mutex);
1326		ret = rdma_set_ib_path(ctx->cm_id, &opa);
1327		mutex_unlock(&ctx->mutex);
1328	} else {
1329		mutex_lock(&ctx->mutex);
1330		ret = rdma_set_ib_path(ctx->cm_id, &sa_path);
1331		mutex_unlock(&ctx->mutex);
1332	}
1333	if (ret)
1334		return ret;
1335
1336	memset(&event, 0, sizeof event);
1337	event.event = RDMA_CM_EVENT_ROUTE_RESOLVED;
1338	return ucma_event_handler(ctx->cm_id, &event);
1339}
1340
1341static int ucma_set_option_ib(struct ucma_context *ctx, int optname,
1342			      void *optval, size_t optlen)
1343{
1344	int ret;
1345
1346	switch (optname) {
1347	case RDMA_OPTION_IB_PATH:
1348		ret = ucma_set_ib_path(ctx, optval, optlen);
1349		break;
1350	default:
1351		ret = -ENOSYS;
1352	}
1353
1354	return ret;
1355}
1356
1357static int ucma_set_option_level(struct ucma_context *ctx, int level,
1358				 int optname, void *optval, size_t optlen)
1359{
1360	int ret;
1361
1362	switch (level) {
1363	case RDMA_OPTION_ID:
1364		mutex_lock(&ctx->mutex);
1365		ret = ucma_set_option_id(ctx, optname, optval, optlen);
1366		mutex_unlock(&ctx->mutex);
1367		break;
1368	case RDMA_OPTION_IB:
1369		ret = ucma_set_option_ib(ctx, optname, optval, optlen);
1370		break;
1371	default:
1372		ret = -ENOSYS;
1373	}
1374
1375	return ret;
1376}
1377
1378static ssize_t ucma_set_option(struct ucma_file *file, const char __user *inbuf,
1379			       int in_len, int out_len)
1380{
1381	struct rdma_ucm_set_option cmd;
1382	struct ucma_context *ctx;
1383	void *optval;
1384	int ret;
1385
1386	if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
1387		return -EFAULT;
1388
1389	if (unlikely(cmd.optlen > KMALLOC_MAX_SIZE))
1390		return -EINVAL;
1391
1392	ctx = ucma_get_ctx(file, cmd.id);
1393	if (IS_ERR(ctx))
1394		return PTR_ERR(ctx);
1395
1396	optval = memdup_user(u64_to_user_ptr(cmd.optval),
1397			     cmd.optlen);
1398	if (IS_ERR(optval)) {
1399		ret = PTR_ERR(optval);
1400		goto out;
1401	}
1402
1403	ret = ucma_set_option_level(ctx, cmd.level, cmd.optname, optval,
1404				    cmd.optlen);
1405	kfree(optval);
1406
1407out:
1408	ucma_put_ctx(ctx);
1409	return ret;
1410}
1411
1412static ssize_t ucma_notify(struct ucma_file *file, const char __user *inbuf,
1413			   int in_len, int out_len)
1414{
1415	struct rdma_ucm_notify cmd;
1416	struct ucma_context *ctx;
1417	int ret = -EINVAL;
1418
1419	if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
1420		return -EFAULT;
1421
1422	ctx = ucma_get_ctx(file, cmd.id);
1423	if (IS_ERR(ctx))
1424		return PTR_ERR(ctx);
1425
1426	mutex_lock(&ctx->mutex);
1427	if (ctx->cm_id->device)
1428		ret = rdma_notify(ctx->cm_id, (enum ib_event_type)cmd.event);
1429	mutex_unlock(&ctx->mutex);
1430
1431	ucma_put_ctx(ctx);
1432	return ret;
1433}
1434
1435static ssize_t ucma_process_join(struct ucma_file *file,
1436				 struct rdma_ucm_join_mcast *cmd,  int out_len)
1437{
1438	struct rdma_ucm_create_id_resp resp;
1439	struct ucma_context *ctx;
1440	struct ucma_multicast *mc;
1441	struct sockaddr *addr;
1442	int ret;
1443	u8 join_state;
1444
1445	if (out_len < sizeof(resp))
1446		return -ENOSPC;
1447
1448	addr = (struct sockaddr *) &cmd->addr;
1449	if (cmd->addr_size != rdma_addr_size(addr))
1450		return -EINVAL;
1451
1452	if (cmd->join_flags == RDMA_MC_JOIN_FLAG_FULLMEMBER)
1453		join_state = BIT(FULLMEMBER_JOIN);
1454	else if (cmd->join_flags == RDMA_MC_JOIN_FLAG_SENDONLY_FULLMEMBER)
1455		join_state = BIT(SENDONLY_FULLMEMBER_JOIN);
1456	else
1457		return -EINVAL;
1458
1459	ctx = ucma_get_ctx_dev(file, cmd->id);
1460	if (IS_ERR(ctx))
1461		return PTR_ERR(ctx);
1462
1463	mc = kzalloc(sizeof(*mc), GFP_KERNEL);
 
1464	if (!mc) {
1465		ret = -ENOMEM;
1466		goto err_put_ctx;
1467	}
1468
1469	mc->ctx = ctx;
1470	mc->join_state = join_state;
1471	mc->uid = cmd->uid;
1472	memcpy(&mc->addr, addr, cmd->addr_size);
1473
1474	xa_lock(&multicast_table);
1475	if (__xa_alloc(&multicast_table, &mc->id, NULL, xa_limit_32b,
1476		     GFP_KERNEL)) {
1477		ret = -ENOMEM;
1478		goto err_free_mc;
1479	}
1480
1481	list_add_tail(&mc->list, &ctx->mc_list);
1482	xa_unlock(&multicast_table);
1483
1484	mutex_lock(&ctx->mutex);
1485	ret = rdma_join_multicast(ctx->cm_id, (struct sockaddr *)&mc->addr,
1486				  join_state, mc);
1487	mutex_unlock(&ctx->mutex);
1488	if (ret)
1489		goto err_xa_erase;
1490
1491	resp.id = mc->id;
1492	if (copy_to_user(u64_to_user_ptr(cmd->response),
1493			 &resp, sizeof(resp))) {
1494		ret = -EFAULT;
1495		goto err_leave_multicast;
1496	}
1497
1498	xa_store(&multicast_table, mc->id, mc, 0);
1499
1500	ucma_put_ctx(ctx);
1501	return 0;
1502
1503err_leave_multicast:
1504	mutex_lock(&ctx->mutex);
1505	rdma_leave_multicast(ctx->cm_id, (struct sockaddr *) &mc->addr);
1506	mutex_unlock(&ctx->mutex);
1507	ucma_cleanup_mc_events(mc);
1508err_xa_erase:
1509	xa_lock(&multicast_table);
 
 
1510	list_del(&mc->list);
1511	__xa_erase(&multicast_table, mc->id);
1512err_free_mc:
1513	xa_unlock(&multicast_table);
1514	kfree(mc);
1515err_put_ctx:
 
1516	ucma_put_ctx(ctx);
1517	return ret;
1518}
1519
1520static ssize_t ucma_join_ip_multicast(struct ucma_file *file,
1521				      const char __user *inbuf,
1522				      int in_len, int out_len)
1523{
1524	struct rdma_ucm_join_ip_mcast cmd;
1525	struct rdma_ucm_join_mcast join_cmd;
1526
1527	if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
1528		return -EFAULT;
1529
1530	join_cmd.response = cmd.response;
1531	join_cmd.uid = cmd.uid;
1532	join_cmd.id = cmd.id;
1533	join_cmd.addr_size = rdma_addr_size_in6(&cmd.addr);
1534	if (!join_cmd.addr_size)
1535		return -EINVAL;
1536
1537	join_cmd.join_flags = RDMA_MC_JOIN_FLAG_FULLMEMBER;
1538	memcpy(&join_cmd.addr, &cmd.addr, join_cmd.addr_size);
1539
1540	return ucma_process_join(file, &join_cmd, out_len);
1541}
1542
1543static ssize_t ucma_join_multicast(struct ucma_file *file,
1544				   const char __user *inbuf,
1545				   int in_len, int out_len)
1546{
1547	struct rdma_ucm_join_mcast cmd;
1548
1549	if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
1550		return -EFAULT;
1551
1552	if (!rdma_addr_size_kss(&cmd.addr))
1553		return -EINVAL;
1554
1555	return ucma_process_join(file, &cmd, out_len);
1556}
1557
1558static ssize_t ucma_leave_multicast(struct ucma_file *file,
1559				    const char __user *inbuf,
1560				    int in_len, int out_len)
1561{
1562	struct rdma_ucm_destroy_id cmd;
1563	struct rdma_ucm_destroy_id_resp resp;
1564	struct ucma_multicast *mc;
1565	int ret = 0;
1566
1567	if (out_len < sizeof(resp))
1568		return -ENOSPC;
1569
1570	if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
1571		return -EFAULT;
1572
1573	xa_lock(&multicast_table);
1574	mc = xa_load(&multicast_table, cmd.id);
1575	if (!mc)
1576		mc = ERR_PTR(-ENOENT);
1577	else if (READ_ONCE(mc->ctx->file) != file)
1578		mc = ERR_PTR(-EINVAL);
1579	else if (!refcount_inc_not_zero(&mc->ctx->ref))
1580		mc = ERR_PTR(-ENXIO);
 
 
 
1581
1582	if (IS_ERR(mc)) {
1583		xa_unlock(&multicast_table);
1584		ret = PTR_ERR(mc);
1585		goto out;
1586	}
1587
1588	list_del(&mc->list);
1589	__xa_erase(&multicast_table, mc->id);
1590	xa_unlock(&multicast_table);
1591
1592	mutex_lock(&mc->ctx->mutex);
1593	rdma_leave_multicast(mc->ctx->cm_id, (struct sockaddr *) &mc->addr);
1594	mutex_unlock(&mc->ctx->mutex);
1595
1596	ucma_cleanup_mc_events(mc);
 
 
1597
1598	ucma_put_ctx(mc->ctx);
1599	resp.events_reported = mc->events_reported;
1600	kfree(mc);
1601
1602	if (copy_to_user(u64_to_user_ptr(cmd.response),
1603			 &resp, sizeof(resp)))
1604		ret = -EFAULT;
1605out:
1606	return ret;
1607}
1608
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1609static ssize_t ucma_migrate_id(struct ucma_file *new_file,
1610			       const char __user *inbuf,
1611			       int in_len, int out_len)
1612{
1613	struct rdma_ucm_migrate_id cmd;
1614	struct rdma_ucm_migrate_resp resp;
1615	struct ucma_event *uevent, *tmp;
1616	struct ucma_context *ctx;
1617	LIST_HEAD(event_list);
1618	struct fd f;
1619	struct ucma_file *cur_file;
1620	int ret = 0;
1621
1622	if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
1623		return -EFAULT;
1624
1625	/* Get current fd to protect against it being closed */
1626	f = fdget(cmd.fd);
1627	if (!f.file)
1628		return -ENOENT;
1629	if (f.file->f_op != &ucma_fops) {
1630		ret = -EINVAL;
1631		goto file_put;
1632	}
1633	cur_file = f.file->private_data;
1634
1635	/* Validate current fd and prevent destruction of id. */
1636	ctx = ucma_get_ctx(cur_file, cmd.id);
1637	if (IS_ERR(ctx)) {
1638		ret = PTR_ERR(ctx);
1639		goto file_put;
1640	}
1641
1642	rdma_lock_handler(ctx->cm_id);
1643	/*
1644	 * ctx->file can only be changed under the handler & xa_lock. xa_load()
1645	 * must be checked again to ensure the ctx hasn't begun destruction
1646	 * since the ucma_get_ctx().
1647	 */
1648	xa_lock(&ctx_table);
1649	if (_ucma_find_context(cmd.id, cur_file) != ctx) {
1650		xa_unlock(&ctx_table);
1651		ret = -ENOENT;
1652		goto err_unlock;
1653	}
1654	ctx->file = new_file;
1655	xa_unlock(&ctx_table);
1656
1657	mutex_lock(&cur_file->mut);
1658	list_del(&ctx->list);
1659	/*
1660	 * At this point lock_handler() prevents addition of new uevents for
1661	 * this ctx.
1662	 */
1663	list_for_each_entry_safe(uevent, tmp, &cur_file->event_list, list)
1664		if (uevent->ctx == ctx)
1665			list_move_tail(&uevent->list, &event_list);
 
 
 
1666	resp.events_reported = ctx->events_reported;
1667	mutex_unlock(&cur_file->mut);
1668
1669	mutex_lock(&new_file->mut);
1670	list_add_tail(&ctx->list, &new_file->ctx_list);
1671	list_splice_tail(&event_list, &new_file->event_list);
1672	mutex_unlock(&new_file->mut);
1673
 
1674	if (copy_to_user(u64_to_user_ptr(cmd.response),
1675			 &resp, sizeof(resp)))
1676		ret = -EFAULT;
1677
1678err_unlock:
1679	rdma_unlock_handler(ctx->cm_id);
1680	ucma_put_ctx(ctx);
1681file_put:
1682	fdput(f);
1683	return ret;
1684}
1685
1686static ssize_t (*ucma_cmd_table[])(struct ucma_file *file,
1687				   const char __user *inbuf,
1688				   int in_len, int out_len) = {
1689	[RDMA_USER_CM_CMD_CREATE_ID] 	 = ucma_create_id,
1690	[RDMA_USER_CM_CMD_DESTROY_ID]	 = ucma_destroy_id,
1691	[RDMA_USER_CM_CMD_BIND_IP]	 = ucma_bind_ip,
1692	[RDMA_USER_CM_CMD_RESOLVE_IP]	 = ucma_resolve_ip,
1693	[RDMA_USER_CM_CMD_RESOLVE_ROUTE] = ucma_resolve_route,
1694	[RDMA_USER_CM_CMD_QUERY_ROUTE]	 = ucma_query_route,
1695	[RDMA_USER_CM_CMD_CONNECT]	 = ucma_connect,
1696	[RDMA_USER_CM_CMD_LISTEN]	 = ucma_listen,
1697	[RDMA_USER_CM_CMD_ACCEPT]	 = ucma_accept,
1698	[RDMA_USER_CM_CMD_REJECT]	 = ucma_reject,
1699	[RDMA_USER_CM_CMD_DISCONNECT]	 = ucma_disconnect,
1700	[RDMA_USER_CM_CMD_INIT_QP_ATTR]	 = ucma_init_qp_attr,
1701	[RDMA_USER_CM_CMD_GET_EVENT]	 = ucma_get_event,
1702	[RDMA_USER_CM_CMD_GET_OPTION]	 = NULL,
1703	[RDMA_USER_CM_CMD_SET_OPTION]	 = ucma_set_option,
1704	[RDMA_USER_CM_CMD_NOTIFY]	 = ucma_notify,
1705	[RDMA_USER_CM_CMD_JOIN_IP_MCAST] = ucma_join_ip_multicast,
1706	[RDMA_USER_CM_CMD_LEAVE_MCAST]	 = ucma_leave_multicast,
1707	[RDMA_USER_CM_CMD_MIGRATE_ID]	 = ucma_migrate_id,
1708	[RDMA_USER_CM_CMD_QUERY]	 = ucma_query,
1709	[RDMA_USER_CM_CMD_BIND]		 = ucma_bind,
1710	[RDMA_USER_CM_CMD_RESOLVE_ADDR]	 = ucma_resolve_addr,
1711	[RDMA_USER_CM_CMD_JOIN_MCAST]	 = ucma_join_multicast
1712};
1713
1714static ssize_t ucma_write(struct file *filp, const char __user *buf,
1715			  size_t len, loff_t *pos)
1716{
1717	struct ucma_file *file = filp->private_data;
1718	struct rdma_ucm_cmd_hdr hdr;
1719	ssize_t ret;
1720
1721	if (!ib_safe_file_access(filp)) {
1722		pr_err_once("%s: process %d (%s) changed security contexts after opening file descriptor, this is not allowed.\n",
1723			    __func__, task_tgid_vnr(current), current->comm);
1724		return -EACCES;
1725	}
1726
1727	if (len < sizeof(hdr))
1728		return -EINVAL;
1729
1730	if (copy_from_user(&hdr, buf, sizeof(hdr)))
1731		return -EFAULT;
1732
1733	if (hdr.cmd >= ARRAY_SIZE(ucma_cmd_table))
1734		return -EINVAL;
1735	hdr.cmd = array_index_nospec(hdr.cmd, ARRAY_SIZE(ucma_cmd_table));
1736
1737	if (hdr.in + sizeof(hdr) > len)
1738		return -EINVAL;
1739
1740	if (!ucma_cmd_table[hdr.cmd])
1741		return -ENOSYS;
1742
1743	ret = ucma_cmd_table[hdr.cmd](file, buf + sizeof(hdr), hdr.in, hdr.out);
1744	if (!ret)
1745		ret = len;
1746
1747	return ret;
1748}
1749
1750static __poll_t ucma_poll(struct file *filp, struct poll_table_struct *wait)
1751{
1752	struct ucma_file *file = filp->private_data;
1753	__poll_t mask = 0;
1754
1755	poll_wait(filp, &file->poll_wait, wait);
1756
1757	if (!list_empty(&file->event_list))
1758		mask = EPOLLIN | EPOLLRDNORM;
1759
1760	return mask;
1761}
1762
1763/*
1764 * ucma_open() does not need the BKL:
1765 *
1766 *  - no global state is referred to;
1767 *  - there is no ioctl method to race against;
1768 *  - no further module initialization is required for open to work
1769 *    after the device is registered.
1770 */
1771static int ucma_open(struct inode *inode, struct file *filp)
1772{
1773	struct ucma_file *file;
1774
1775	file = kmalloc(sizeof *file, GFP_KERNEL);
1776	if (!file)
1777		return -ENOMEM;
1778
 
 
 
 
 
 
 
1779	INIT_LIST_HEAD(&file->event_list);
1780	INIT_LIST_HEAD(&file->ctx_list);
1781	init_waitqueue_head(&file->poll_wait);
1782	mutex_init(&file->mut);
1783
1784	filp->private_data = file;
1785	file->filp = filp;
1786
1787	return stream_open(inode, filp);
1788}
1789
1790static int ucma_close(struct inode *inode, struct file *filp)
1791{
1792	struct ucma_file *file = filp->private_data;
 
1793
1794	/*
1795	 * All paths that touch ctx_list or ctx_list starting from write() are
1796	 * prevented by this being a FD release function. The list_add_tail() in
1797	 * ucma_connect_event_handler() can run concurrently, however it only
1798	 * adds to the list *after* a listening ID. By only reading the first of
1799	 * the list, and relying on ucma_destroy_private_ctx() to block
1800	 * ucma_connect_event_handler(), no additional locking is needed.
1801	 */
1802	while (!list_empty(&file->ctx_list)) {
1803		struct ucma_context *ctx = list_first_entry(
1804			&file->ctx_list, struct ucma_context, list);
1805
1806		WARN_ON(xa_cmpxchg(&ctx_table, ctx->id, ctx, XA_ZERO_ENTRY,
1807				   GFP_KERNEL) != ctx);
1808		ucma_destroy_private_ctx(ctx);
 
 
 
 
 
 
 
 
 
 
 
 
1809	}
 
 
1810	kfree(file);
1811	return 0;
1812}
1813
1814static const struct file_operations ucma_fops = {
1815	.owner 	 = THIS_MODULE,
1816	.open 	 = ucma_open,
1817	.release = ucma_close,
1818	.write	 = ucma_write,
1819	.poll    = ucma_poll,
1820	.llseek	 = no_llseek,
1821};
1822
1823static struct miscdevice ucma_misc = {
1824	.minor		= MISC_DYNAMIC_MINOR,
1825	.name		= "rdma_cm",
1826	.nodename	= "infiniband/rdma_cm",
1827	.mode		= 0666,
1828	.fops		= &ucma_fops,
1829};
1830
1831static int ucma_get_global_nl_info(struct ib_client_nl_info *res)
1832{
1833	res->abi = RDMA_USER_CM_ABI_VERSION;
1834	res->cdev = ucma_misc.this_device;
1835	return 0;
1836}
1837
1838static struct ib_client rdma_cma_client = {
1839	.name = "rdma_cm",
1840	.get_global_nl_info = ucma_get_global_nl_info,
1841};
1842MODULE_ALIAS_RDMA_CLIENT("rdma_cm");
1843
1844static ssize_t abi_version_show(struct device *dev,
1845				struct device_attribute *attr, char *buf)
1846{
1847	return sysfs_emit(buf, "%d\n", RDMA_USER_CM_ABI_VERSION);
1848}
1849static DEVICE_ATTR_RO(abi_version);
1850
1851static int __init ucma_init(void)
1852{
1853	int ret;
1854
1855	ret = misc_register(&ucma_misc);
1856	if (ret)
1857		return ret;
1858
1859	ret = device_create_file(ucma_misc.this_device, &dev_attr_abi_version);
1860	if (ret) {
1861		pr_err("rdma_ucm: couldn't create abi_version attr\n");
1862		goto err1;
1863	}
1864
1865	ucma_ctl_table_hdr = register_net_sysctl(&init_net, "net/rdma_ucm", ucma_ctl_table);
1866	if (!ucma_ctl_table_hdr) {
1867		pr_err("rdma_ucm: couldn't register sysctl paths\n");
1868		ret = -ENOMEM;
1869		goto err2;
1870	}
1871
1872	ret = ib_register_client(&rdma_cma_client);
1873	if (ret)
1874		goto err3;
1875
1876	return 0;
1877err3:
1878	unregister_net_sysctl_table(ucma_ctl_table_hdr);
1879err2:
1880	device_remove_file(ucma_misc.this_device, &dev_attr_abi_version);
1881err1:
1882	misc_deregister(&ucma_misc);
1883	return ret;
1884}
1885
1886static void __exit ucma_cleanup(void)
1887{
1888	ib_unregister_client(&rdma_cma_client);
1889	unregister_net_sysctl_table(ucma_ctl_table_hdr);
1890	device_remove_file(ucma_misc.this_device, &dev_attr_abi_version);
1891	misc_deregister(&ucma_misc);
 
 
1892}
1893
1894module_init(ucma_init);
1895module_exit(ucma_cleanup);
v4.17
   1/*
   2 * Copyright (c) 2005-2006 Intel Corporation.  All rights reserved.
   3 *
   4 * This software is available to you under a choice of one of two
   5 * licenses.  You may choose to be licensed under the terms of the GNU
   6 * General Public License (GPL) Version 2, available from the file
   7 * COPYING in the main directory of this source tree, or the
   8 * OpenIB.org BSD license below:
   9 *
  10 *     Redistribution and use in source and binary forms, with or
  11 *     without modification, are permitted provided that the following
  12 *     conditions are met:
  13 *
  14 *      - Redistributions of source code must retain the above
  15 *	copyright notice, this list of conditions and the following
  16 *	disclaimer.
  17 *
  18 *      - Redistributions in binary form must reproduce the above
  19 *	copyright notice, this list of conditions and the following
  20 *	disclaimer in the documentation and/or other materials
  21 *	provided with the distribution.
  22 *
  23 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
  24 * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
  25 * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
  26 * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
  27 * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
  28 * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
  29 * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
  30 * SOFTWARE.
  31 */
  32
  33#include <linux/completion.h>
  34#include <linux/file.h>
  35#include <linux/mutex.h>
  36#include <linux/poll.h>
  37#include <linux/sched.h>
  38#include <linux/idr.h>
  39#include <linux/in.h>
  40#include <linux/in6.h>
  41#include <linux/miscdevice.h>
  42#include <linux/slab.h>
  43#include <linux/sysctl.h>
  44#include <linux/module.h>
  45#include <linux/nsproxy.h>
  46
 
 
  47#include <rdma/rdma_user_cm.h>
  48#include <rdma/ib_marshall.h>
  49#include <rdma/rdma_cm.h>
  50#include <rdma/rdma_cm_ib.h>
  51#include <rdma/ib_addr.h>
  52#include <rdma/ib.h>
 
 
 
  53
  54MODULE_AUTHOR("Sean Hefty");
  55MODULE_DESCRIPTION("RDMA Userspace Connection Manager Access");
  56MODULE_LICENSE("Dual BSD/GPL");
  57
  58static unsigned int max_backlog = 1024;
  59
  60static struct ctl_table_header *ucma_ctl_table_hdr;
  61static struct ctl_table ucma_ctl_table[] = {
  62	{
  63		.procname	= "max_backlog",
  64		.data		= &max_backlog,
  65		.maxlen		= sizeof max_backlog,
  66		.mode		= 0644,
  67		.proc_handler	= proc_dointvec,
  68	},
  69	{ }
  70};
  71
  72struct ucma_file {
  73	struct mutex		mut;
  74	struct file		*filp;
  75	struct list_head	ctx_list;
  76	struct list_head	event_list;
  77	wait_queue_head_t	poll_wait;
  78	struct workqueue_struct	*close_wq;
  79};
  80
  81struct ucma_context {
  82	int			id;
  83	struct completion	comp;
  84	atomic_t		ref;
  85	int			events_reported;
  86	int			backlog;
  87
  88	struct ucma_file	*file;
  89	struct rdma_cm_id	*cm_id;
 
  90	u64			uid;
  91
  92	struct list_head	list;
  93	struct list_head	mc_list;
  94	/* mark that device is in process of destroying the internal HW
  95	 * resources, protected by the global mut
  96	 */
  97	int			closing;
  98	/* sync between removal event and id destroy, protected by file mut */
  99	int			destroying;
 100	struct work_struct	close_work;
 101};
 102
 103struct ucma_multicast {
 104	struct ucma_context	*ctx;
 105	int			id;
 106	int			events_reported;
 107
 108	u64			uid;
 109	u8			join_state;
 110	struct list_head	list;
 111	struct sockaddr_storage	addr;
 112};
 113
 114struct ucma_event {
 115	struct ucma_context	*ctx;
 
 116	struct ucma_multicast	*mc;
 117	struct list_head	list;
 118	struct rdma_cm_id	*cm_id;
 119	struct rdma_ucm_event_resp resp;
 120	struct work_struct	close_work;
 121};
 122
 123static DEFINE_MUTEX(mut);
 124static DEFINE_IDR(ctx_idr);
 125static DEFINE_IDR(multicast_idr);
 
 
 126
 127static inline struct ucma_context *_ucma_find_context(int id,
 128						      struct ucma_file *file)
 129{
 130	struct ucma_context *ctx;
 131
 132	ctx = idr_find(&ctx_idr, id);
 133	if (!ctx)
 134		ctx = ERR_PTR(-ENOENT);
 135	else if (ctx->file != file || !ctx->cm_id)
 136		ctx = ERR_PTR(-EINVAL);
 137	return ctx;
 138}
 139
 140static struct ucma_context *ucma_get_ctx(struct ucma_file *file, int id)
 141{
 142	struct ucma_context *ctx;
 143
 144	mutex_lock(&mut);
 145	ctx = _ucma_find_context(id, file);
 146	if (!IS_ERR(ctx)) {
 147		if (ctx->closing)
 148			ctx = ERR_PTR(-EIO);
 149		else
 150			atomic_inc(&ctx->ref);
 151	}
 152	mutex_unlock(&mut);
 153	return ctx;
 154}
 155
 156static void ucma_put_ctx(struct ucma_context *ctx)
 157{
 158	if (atomic_dec_and_test(&ctx->ref))
 159		complete(&ctx->comp);
 160}
 161
 162/*
 163 * Same as ucm_get_ctx but requires that ->cm_id->device is valid, eg that the
 164 * CM_ID is bound.
 165 */
 166static struct ucma_context *ucma_get_ctx_dev(struct ucma_file *file, int id)
 167{
 168	struct ucma_context *ctx = ucma_get_ctx(file, id);
 169
 170	if (IS_ERR(ctx))
 171		return ctx;
 172	if (!ctx->cm_id->device) {
 173		ucma_put_ctx(ctx);
 174		return ERR_PTR(-EINVAL);
 175	}
 176	return ctx;
 177}
 178
 179static void ucma_close_event_id(struct work_struct *work)
 180{
 181	struct ucma_event *uevent_close =  container_of(work, struct ucma_event, close_work);
 182
 183	rdma_destroy_id(uevent_close->cm_id);
 184	kfree(uevent_close);
 185}
 186
 187static void ucma_close_id(struct work_struct *work)
 188{
 189	struct ucma_context *ctx =  container_of(work, struct ucma_context, close_work);
 190
 191	/* once all inflight tasks are finished, we close all underlying
 192	 * resources. The context is still alive till its explicit destryoing
 193	 * by its creator.
 194	 */
 195	ucma_put_ctx(ctx);
 196	wait_for_completion(&ctx->comp);
 197	/* No new events will be generated after destroying the id. */
 198	rdma_destroy_id(ctx->cm_id);
 
 
 
 199}
 200
 201static struct ucma_context *ucma_alloc_ctx(struct ucma_file *file)
 202{
 203	struct ucma_context *ctx;
 204
 205	ctx = kzalloc(sizeof(*ctx), GFP_KERNEL);
 206	if (!ctx)
 207		return NULL;
 208
 209	INIT_WORK(&ctx->close_work, ucma_close_id);
 210	atomic_set(&ctx->ref, 1);
 211	init_completion(&ctx->comp);
 212	INIT_LIST_HEAD(&ctx->mc_list);
 
 
 213	ctx->file = file;
 
 214
 215	mutex_lock(&mut);
 216	ctx->id = idr_alloc(&ctx_idr, ctx, 0, 0, GFP_KERNEL);
 217	mutex_unlock(&mut);
 218	if (ctx->id < 0)
 219		goto error;
 220
 221	list_add_tail(&ctx->list, &file->ctx_list);
 222	return ctx;
 
 223
 224error:
 225	kfree(ctx);
 226	return NULL;
 
 
 227}
 228
 229static struct ucma_multicast* ucma_alloc_multicast(struct ucma_context *ctx)
 230{
 231	struct ucma_multicast *mc;
 232
 233	mc = kzalloc(sizeof(*mc), GFP_KERNEL);
 234	if (!mc)
 235		return NULL;
 236
 237	mutex_lock(&mut);
 238	mc->id = idr_alloc(&multicast_idr, mc, 0, 0, GFP_KERNEL);
 239	mutex_unlock(&mut);
 240	if (mc->id < 0)
 241		goto error;
 242
 243	mc->ctx = ctx;
 244	list_add_tail(&mc->list, &ctx->mc_list);
 245	return mc;
 246
 247error:
 248	kfree(mc);
 249	return NULL;
 250}
 251
 252static void ucma_copy_conn_event(struct rdma_ucm_conn_param *dst,
 253				 struct rdma_conn_param *src)
 254{
 255	if (src->private_data_len)
 256		memcpy(dst->private_data, src->private_data,
 257		       src->private_data_len);
 258	dst->private_data_len = src->private_data_len;
 259	dst->responder_resources =src->responder_resources;
 260	dst->initiator_depth = src->initiator_depth;
 261	dst->flow_control = src->flow_control;
 262	dst->retry_count = src->retry_count;
 263	dst->rnr_retry_count = src->rnr_retry_count;
 264	dst->srq = src->srq;
 265	dst->qp_num = src->qp_num;
 266}
 267
 268static void ucma_copy_ud_event(struct ib_device *device,
 269			       struct rdma_ucm_ud_param *dst,
 270			       struct rdma_ud_param *src)
 271{
 272	if (src->private_data_len)
 273		memcpy(dst->private_data, src->private_data,
 274		       src->private_data_len);
 275	dst->private_data_len = src->private_data_len;
 276	ib_copy_ah_attr_to_user(device, &dst->ah_attr, &src->ah_attr);
 277	dst->qp_num = src->qp_num;
 278	dst->qkey = src->qkey;
 279}
 280
 281static void ucma_set_event_context(struct ucma_context *ctx,
 282				   struct rdma_cm_event *event,
 283				   struct ucma_event *uevent)
 284{
 
 
 
 
 
 
 285	uevent->ctx = ctx;
 286	switch (event->event) {
 287	case RDMA_CM_EVENT_MULTICAST_JOIN:
 288	case RDMA_CM_EVENT_MULTICAST_ERROR:
 289		uevent->mc = (struct ucma_multicast *)
 290			     event->param.ud.private_data;
 291		uevent->resp.uid = uevent->mc->uid;
 292		uevent->resp.id = uevent->mc->id;
 293		break;
 294	default:
 295		uevent->resp.uid = ctx->uid;
 296		uevent->resp.id = ctx->id;
 297		break;
 298	}
 
 
 
 
 
 
 
 
 
 
 
 
 299}
 300
 301/* Called with file->mut locked for the relevant context. */
 302static void ucma_removal_event_handler(struct rdma_cm_id *cm_id)
 303{
 304	struct ucma_context *ctx = cm_id->context;
 305	struct ucma_event *con_req_eve;
 306	int event_found = 0;
 
 
 
 
 
 
 
 
 
 
 
 
 
 307
 308	if (ctx->destroying)
 309		return;
 310
 311	/* only if context is pointing to cm_id that it owns it and can be
 312	 * queued to be closed, otherwise that cm_id is an inflight one that
 313	 * is part of that context event list pending to be detached and
 314	 * reattached to its new context as part of ucma_get_event,
 315	 * handled separately below.
 316	 */
 317	if (ctx->cm_id == cm_id) {
 318		mutex_lock(&mut);
 319		ctx->closing = 1;
 320		mutex_unlock(&mut);
 321		queue_work(ctx->file->close_wq, &ctx->close_work);
 322		return;
 323	}
 324
 325	list_for_each_entry(con_req_eve, &ctx->file->event_list, list) {
 326		if (con_req_eve->cm_id == cm_id &&
 327		    con_req_eve->resp.event == RDMA_CM_EVENT_CONNECT_REQUEST) {
 328			list_del(&con_req_eve->list);
 329			INIT_WORK(&con_req_eve->close_work, ucma_close_event_id);
 330			queue_work(ctx->file->close_wq, &con_req_eve->close_work);
 331			event_found = 1;
 332			break;
 333		}
 334	}
 335	if (!event_found)
 336		pr_err("ucma_removal_event_handler: warning: connect request event wasn't found\n");
 337}
 338
 339static int ucma_event_handler(struct rdma_cm_id *cm_id,
 340			      struct rdma_cm_event *event)
 341{
 342	struct ucma_event *uevent;
 343	struct ucma_context *ctx = cm_id->context;
 344	int ret = 0;
 345
 346	uevent = kzalloc(sizeof(*uevent), GFP_KERNEL);
 347	if (!uevent)
 348		return event->event == RDMA_CM_EVENT_CONNECT_REQUEST;
 349
 350	mutex_lock(&ctx->file->mut);
 351	uevent->cm_id = cm_id;
 352	ucma_set_event_context(ctx, event, uevent);
 353	uevent->resp.event = event->event;
 354	uevent->resp.status = event->status;
 355	if (cm_id->qp_type == IB_QPT_UD)
 356		ucma_copy_ud_event(cm_id->device, &uevent->resp.param.ud,
 357				   &event->param.ud);
 358	else
 359		ucma_copy_conn_event(&uevent->resp.param.conn,
 360				     &event->param.conn);
 361
 362	if (event->event == RDMA_CM_EVENT_CONNECT_REQUEST) {
 363		if (!ctx->backlog) {
 364			ret = -ENOMEM;
 365			kfree(uevent);
 366			goto out;
 367		}
 368		ctx->backlog--;
 369	} else if (!ctx->uid || ctx->cm_id != cm_id) {
 370		/*
 371		 * We ignore events for new connections until userspace has set
 372		 * their context.  This can only happen if an error occurs on a
 373		 * new connection before the user accepts it.  This is okay,
 374		 * since the accept will just fail later. However, we do need
 375		 * to release the underlying HW resources in case of a device
 376		 * removal event.
 377		 */
 378		if (event->event == RDMA_CM_EVENT_DEVICE_REMOVAL)
 379			ucma_removal_event_handler(cm_id);
 380
 381		kfree(uevent);
 382		goto out;
 383	}
 384
 385	list_add_tail(&uevent->list, &ctx->file->event_list);
 386	wake_up_interruptible(&ctx->file->poll_wait);
 387	if (event->event == RDMA_CM_EVENT_DEVICE_REMOVAL)
 388		ucma_removal_event_handler(cm_id);
 389out:
 390	mutex_unlock(&ctx->file->mut);
 391	return ret;
 392}
 393
 394static ssize_t ucma_get_event(struct ucma_file *file, const char __user *inbuf,
 395			      int in_len, int out_len)
 396{
 397	struct ucma_context *ctx;
 398	struct rdma_ucm_get_event cmd;
 399	struct ucma_event *uevent;
 400	int ret = 0;
 401
 402	/*
 403	 * Old 32 bit user space does not send the 4 byte padding in the
 404	 * reserved field. We don't care, allow it to keep working.
 405	 */
 406	if (out_len < sizeof(uevent->resp) - sizeof(uevent->resp.reserved))
 
 407		return -ENOSPC;
 408
 409	if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
 410		return -EFAULT;
 411
 412	mutex_lock(&file->mut);
 413	while (list_empty(&file->event_list)) {
 414		mutex_unlock(&file->mut);
 415
 416		if (file->filp->f_flags & O_NONBLOCK)
 417			return -EAGAIN;
 418
 419		if (wait_event_interruptible(file->poll_wait,
 420					     !list_empty(&file->event_list)))
 421			return -ERESTARTSYS;
 422
 423		mutex_lock(&file->mut);
 424	}
 425
 426	uevent = list_entry(file->event_list.next, struct ucma_event, list);
 427
 428	if (uevent->resp.event == RDMA_CM_EVENT_CONNECT_REQUEST) {
 429		ctx = ucma_alloc_ctx(file);
 430		if (!ctx) {
 431			ret = -ENOMEM;
 432			goto done;
 433		}
 434		uevent->ctx->backlog++;
 435		ctx->cm_id = uevent->cm_id;
 436		ctx->cm_id->context = ctx;
 437		uevent->resp.id = ctx->id;
 438	}
 439
 440	if (copy_to_user(u64_to_user_ptr(cmd.response),
 441			 &uevent->resp,
 442			 min_t(size_t, out_len, sizeof(uevent->resp)))) {
 443		ret = -EFAULT;
 444		goto done;
 445	}
 446
 447	list_del(&uevent->list);
 448	uevent->ctx->events_reported++;
 449	if (uevent->mc)
 450		uevent->mc->events_reported++;
 
 
 
 
 451	kfree(uevent);
 452done:
 453	mutex_unlock(&file->mut);
 454	return ret;
 455}
 456
 457static int ucma_get_qp_type(struct rdma_ucm_create_id *cmd, enum ib_qp_type *qp_type)
 458{
 459	switch (cmd->ps) {
 460	case RDMA_PS_TCP:
 461		*qp_type = IB_QPT_RC;
 462		return 0;
 463	case RDMA_PS_UDP:
 464	case RDMA_PS_IPOIB:
 465		*qp_type = IB_QPT_UD;
 466		return 0;
 467	case RDMA_PS_IB:
 468		*qp_type = cmd->qp_type;
 469		return 0;
 470	default:
 471		return -EINVAL;
 472	}
 473}
 474
 475static ssize_t ucma_create_id(struct ucma_file *file, const char __user *inbuf,
 476			      int in_len, int out_len)
 477{
 478	struct rdma_ucm_create_id cmd;
 479	struct rdma_ucm_create_id_resp resp;
 480	struct ucma_context *ctx;
 481	struct rdma_cm_id *cm_id;
 482	enum ib_qp_type qp_type;
 483	int ret;
 484
 485	if (out_len < sizeof(resp))
 486		return -ENOSPC;
 487
 488	if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
 489		return -EFAULT;
 490
 491	ret = ucma_get_qp_type(&cmd, &qp_type);
 492	if (ret)
 493		return ret;
 494
 495	mutex_lock(&file->mut);
 496	ctx = ucma_alloc_ctx(file);
 497	mutex_unlock(&file->mut);
 498	if (!ctx)
 499		return -ENOMEM;
 500
 501	ctx->uid = cmd.uid;
 502	cm_id = __rdma_create_id(current->nsproxy->net_ns,
 503				 ucma_event_handler, ctx, cmd.ps, qp_type, NULL);
 504	if (IS_ERR(cm_id)) {
 505		ret = PTR_ERR(cm_id);
 506		goto err1;
 507	}
 
 508
 509	resp.id = ctx->id;
 510	if (copy_to_user(u64_to_user_ptr(cmd.response),
 511			 &resp, sizeof(resp))) {
 512		ret = -EFAULT;
 513		goto err2;
 514	}
 515
 516	ctx->cm_id = cm_id;
 
 
 517	return 0;
 518
 519err2:
 520	rdma_destroy_id(cm_id);
 521err1:
 522	mutex_lock(&mut);
 523	idr_remove(&ctx_idr, ctx->id);
 524	mutex_unlock(&mut);
 525	mutex_lock(&file->mut);
 526	list_del(&ctx->list);
 527	mutex_unlock(&file->mut);
 528	kfree(ctx);
 529	return ret;
 530}
 531
 532static void ucma_cleanup_multicast(struct ucma_context *ctx)
 533{
 534	struct ucma_multicast *mc, *tmp;
 535
 536	mutex_lock(&mut);
 537	list_for_each_entry_safe(mc, tmp, &ctx->mc_list, list) {
 538		list_del(&mc->list);
 539		idr_remove(&multicast_idr, mc->id);
 
 
 
 
 540		kfree(mc);
 541	}
 542	mutex_unlock(&mut);
 543}
 544
 545static void ucma_cleanup_mc_events(struct ucma_multicast *mc)
 546{
 547	struct ucma_event *uevent, *tmp;
 548
 
 
 549	list_for_each_entry_safe(uevent, tmp, &mc->ctx->file->event_list, list) {
 550		if (uevent->mc != mc)
 551			continue;
 552
 553		list_del(&uevent->list);
 554		kfree(uevent);
 555	}
 
 
 556}
 557
 558/*
 559 * ucma_free_ctx is called after the underlying rdma CM-ID is destroyed. At
 560 * this point, no new events will be reported from the hardware. However, we
 561 * still need to cleanup the UCMA context for this ID. Specifically, there
 562 * might be events that have not yet been consumed by the user space software.
 563 * These might include pending connect requests which we have not completed
 564 * processing.  We cannot call rdma_destroy_id while holding the lock of the
 565 * context (file->mut), as it might cause a deadlock. We therefore extract all
 566 * relevant events from the context pending events list while holding the
 567 * mutex. After that we release them as needed.
 568 */
 569static int ucma_free_ctx(struct ucma_context *ctx)
 570{
 571	int events_reported;
 572	struct ucma_event *uevent, *tmp;
 573	LIST_HEAD(list);
 574
 575
 576	ucma_cleanup_multicast(ctx);
 577
 578	/* Cleanup events not yet reported to the user. */
 579	mutex_lock(&ctx->file->mut);
 580	list_for_each_entry_safe(uevent, tmp, &ctx->file->event_list, list) {
 581		if (uevent->ctx == ctx)
 
 
 
 
 
 
 582			list_move_tail(&uevent->list, &list);
 
 
 
 
 583	}
 584	list_del(&ctx->list);
 
 585	mutex_unlock(&ctx->file->mut);
 586
 
 
 
 
 
 587	list_for_each_entry_safe(uevent, tmp, &list, list) {
 588		list_del(&uevent->list);
 589		if (uevent->resp.event == RDMA_CM_EVENT_CONNECT_REQUEST)
 590			rdma_destroy_id(uevent->cm_id);
 591		kfree(uevent);
 592	}
 
 
 593
 594	events_reported = ctx->events_reported;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 595	kfree(ctx);
 596	return events_reported;
 597}
 598
 599static ssize_t ucma_destroy_id(struct ucma_file *file, const char __user *inbuf,
 600			       int in_len, int out_len)
 601{
 602	struct rdma_ucm_destroy_id cmd;
 603	struct rdma_ucm_destroy_id_resp resp;
 604	struct ucma_context *ctx;
 605	int ret = 0;
 606
 607	if (out_len < sizeof(resp))
 608		return -ENOSPC;
 609
 610	if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
 611		return -EFAULT;
 612
 613	mutex_lock(&mut);
 614	ctx = _ucma_find_context(cmd.id, file);
 615	if (!IS_ERR(ctx))
 616		idr_remove(&ctx_idr, ctx->id);
 617	mutex_unlock(&mut);
 
 
 
 618
 619	if (IS_ERR(ctx))
 620		return PTR_ERR(ctx);
 621
 622	mutex_lock(&ctx->file->mut);
 623	ctx->destroying = 1;
 624	mutex_unlock(&ctx->file->mut);
 625
 626	flush_workqueue(ctx->file->close_wq);
 627	/* At this point it's guaranteed that there is no inflight
 628	 * closing task */
 629	mutex_lock(&mut);
 630	if (!ctx->closing) {
 631		mutex_unlock(&mut);
 632		ucma_put_ctx(ctx);
 633		wait_for_completion(&ctx->comp);
 634		rdma_destroy_id(ctx->cm_id);
 635	} else {
 636		mutex_unlock(&mut);
 637	}
 638
 639	resp.events_reported = ucma_free_ctx(ctx);
 640	if (copy_to_user(u64_to_user_ptr(cmd.response),
 641			 &resp, sizeof(resp)))
 642		ret = -EFAULT;
 643
 644	return ret;
 645}
 646
 647static ssize_t ucma_bind_ip(struct ucma_file *file, const char __user *inbuf,
 648			      int in_len, int out_len)
 649{
 650	struct rdma_ucm_bind_ip cmd;
 651	struct ucma_context *ctx;
 652	int ret;
 653
 654	if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
 655		return -EFAULT;
 656
 657	if (!rdma_addr_size_in6(&cmd.addr))
 658		return -EINVAL;
 659
 660	ctx = ucma_get_ctx(file, cmd.id);
 661	if (IS_ERR(ctx))
 662		return PTR_ERR(ctx);
 663
 
 664	ret = rdma_bind_addr(ctx->cm_id, (struct sockaddr *) &cmd.addr);
 
 
 665	ucma_put_ctx(ctx);
 666	return ret;
 667}
 668
 669static ssize_t ucma_bind(struct ucma_file *file, const char __user *inbuf,
 670			 int in_len, int out_len)
 671{
 672	struct rdma_ucm_bind cmd;
 673	struct ucma_context *ctx;
 674	int ret;
 675
 676	if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
 677		return -EFAULT;
 678
 679	if (cmd.reserved || !cmd.addr_size ||
 680	    cmd.addr_size != rdma_addr_size_kss(&cmd.addr))
 681		return -EINVAL;
 682
 683	ctx = ucma_get_ctx(file, cmd.id);
 684	if (IS_ERR(ctx))
 685		return PTR_ERR(ctx);
 686
 
 687	ret = rdma_bind_addr(ctx->cm_id, (struct sockaddr *) &cmd.addr);
 
 688	ucma_put_ctx(ctx);
 689	return ret;
 690}
 691
 692static ssize_t ucma_resolve_ip(struct ucma_file *file,
 693			       const char __user *inbuf,
 694			       int in_len, int out_len)
 695{
 696	struct rdma_ucm_resolve_ip cmd;
 697	struct ucma_context *ctx;
 698	int ret;
 699
 700	if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
 701		return -EFAULT;
 702
 703	if ((cmd.src_addr.sin6_family && !rdma_addr_size_in6(&cmd.src_addr)) ||
 704	    !rdma_addr_size_in6(&cmd.dst_addr))
 705		return -EINVAL;
 706
 707	ctx = ucma_get_ctx(file, cmd.id);
 708	if (IS_ERR(ctx))
 709		return PTR_ERR(ctx);
 710
 
 711	ret = rdma_resolve_addr(ctx->cm_id, (struct sockaddr *) &cmd.src_addr,
 712				(struct sockaddr *) &cmd.dst_addr, cmd.timeout_ms);
 
 713	ucma_put_ctx(ctx);
 714	return ret;
 715}
 716
 717static ssize_t ucma_resolve_addr(struct ucma_file *file,
 718				 const char __user *inbuf,
 719				 int in_len, int out_len)
 720{
 721	struct rdma_ucm_resolve_addr cmd;
 722	struct ucma_context *ctx;
 723	int ret;
 724
 725	if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
 726		return -EFAULT;
 727
 728	if (cmd.reserved ||
 729	    (cmd.src_size && (cmd.src_size != rdma_addr_size_kss(&cmd.src_addr))) ||
 730	    !cmd.dst_size || (cmd.dst_size != rdma_addr_size_kss(&cmd.dst_addr)))
 731		return -EINVAL;
 732
 733	ctx = ucma_get_ctx(file, cmd.id);
 734	if (IS_ERR(ctx))
 735		return PTR_ERR(ctx);
 736
 
 737	ret = rdma_resolve_addr(ctx->cm_id, (struct sockaddr *) &cmd.src_addr,
 738				(struct sockaddr *) &cmd.dst_addr, cmd.timeout_ms);
 
 739	ucma_put_ctx(ctx);
 740	return ret;
 741}
 742
 743static ssize_t ucma_resolve_route(struct ucma_file *file,
 744				  const char __user *inbuf,
 745				  int in_len, int out_len)
 746{
 747	struct rdma_ucm_resolve_route cmd;
 748	struct ucma_context *ctx;
 749	int ret;
 750
 751	if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
 752		return -EFAULT;
 753
 754	ctx = ucma_get_ctx_dev(file, cmd.id);
 755	if (IS_ERR(ctx))
 756		return PTR_ERR(ctx);
 757
 
 758	ret = rdma_resolve_route(ctx->cm_id, cmd.timeout_ms);
 
 759	ucma_put_ctx(ctx);
 760	return ret;
 761}
 762
 763static void ucma_copy_ib_route(struct rdma_ucm_query_route_resp *resp,
 764			       struct rdma_route *route)
 765{
 766	struct rdma_dev_addr *dev_addr;
 767
 768	resp->num_paths = route->num_paths;
 769	switch (route->num_paths) {
 770	case 0:
 771		dev_addr = &route->addr.dev_addr;
 772		rdma_addr_get_dgid(dev_addr,
 773				   (union ib_gid *) &resp->ib_route[0].dgid);
 774		rdma_addr_get_sgid(dev_addr,
 775				   (union ib_gid *) &resp->ib_route[0].sgid);
 776		resp->ib_route[0].pkey = cpu_to_be16(ib_addr_get_pkey(dev_addr));
 777		break;
 778	case 2:
 779		ib_copy_path_rec_to_user(&resp->ib_route[1],
 780					 &route->path_rec[1]);
 781		/* fall through */
 782	case 1:
 783		ib_copy_path_rec_to_user(&resp->ib_route[0],
 784					 &route->path_rec[0]);
 785		break;
 786	default:
 787		break;
 788	}
 789}
 790
 791static void ucma_copy_iboe_route(struct rdma_ucm_query_route_resp *resp,
 792				 struct rdma_route *route)
 793{
 794
 795	resp->num_paths = route->num_paths;
 796	switch (route->num_paths) {
 797	case 0:
 798		rdma_ip2gid((struct sockaddr *)&route->addr.dst_addr,
 799			    (union ib_gid *)&resp->ib_route[0].dgid);
 800		rdma_ip2gid((struct sockaddr *)&route->addr.src_addr,
 801			    (union ib_gid *)&resp->ib_route[0].sgid);
 802		resp->ib_route[0].pkey = cpu_to_be16(0xffff);
 803		break;
 804	case 2:
 805		ib_copy_path_rec_to_user(&resp->ib_route[1],
 806					 &route->path_rec[1]);
 807		/* fall through */
 808	case 1:
 809		ib_copy_path_rec_to_user(&resp->ib_route[0],
 810					 &route->path_rec[0]);
 811		break;
 812	default:
 813		break;
 814	}
 815}
 816
 817static void ucma_copy_iw_route(struct rdma_ucm_query_route_resp *resp,
 818			       struct rdma_route *route)
 819{
 820	struct rdma_dev_addr *dev_addr;
 821
 822	dev_addr = &route->addr.dev_addr;
 823	rdma_addr_get_dgid(dev_addr, (union ib_gid *) &resp->ib_route[0].dgid);
 824	rdma_addr_get_sgid(dev_addr, (union ib_gid *) &resp->ib_route[0].sgid);
 825}
 826
 827static ssize_t ucma_query_route(struct ucma_file *file,
 828				const char __user *inbuf,
 829				int in_len, int out_len)
 830{
 831	struct rdma_ucm_query cmd;
 832	struct rdma_ucm_query_route_resp resp;
 833	struct ucma_context *ctx;
 834	struct sockaddr *addr;
 835	int ret = 0;
 836
 837	if (out_len < sizeof(resp))
 838		return -ENOSPC;
 839
 840	if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
 841		return -EFAULT;
 842
 843	ctx = ucma_get_ctx(file, cmd.id);
 844	if (IS_ERR(ctx))
 845		return PTR_ERR(ctx);
 846
 
 847	memset(&resp, 0, sizeof resp);
 848	addr = (struct sockaddr *) &ctx->cm_id->route.addr.src_addr;
 849	memcpy(&resp.src_addr, addr, addr->sa_family == AF_INET ?
 850				     sizeof(struct sockaddr_in) :
 851				     sizeof(struct sockaddr_in6));
 852	addr = (struct sockaddr *) &ctx->cm_id->route.addr.dst_addr;
 853	memcpy(&resp.dst_addr, addr, addr->sa_family == AF_INET ?
 854				     sizeof(struct sockaddr_in) :
 855				     sizeof(struct sockaddr_in6));
 856	if (!ctx->cm_id->device)
 857		goto out;
 858
 859	resp.node_guid = (__force __u64) ctx->cm_id->device->node_guid;
 
 860	resp.port_num = ctx->cm_id->port_num;
 861
 862	if (rdma_cap_ib_sa(ctx->cm_id->device, ctx->cm_id->port_num))
 863		ucma_copy_ib_route(&resp, &ctx->cm_id->route);
 864	else if (rdma_protocol_roce(ctx->cm_id->device, ctx->cm_id->port_num))
 865		ucma_copy_iboe_route(&resp, &ctx->cm_id->route);
 866	else if (rdma_protocol_iwarp(ctx->cm_id->device, ctx->cm_id->port_num))
 867		ucma_copy_iw_route(&resp, &ctx->cm_id->route);
 868
 869out:
 870	if (copy_to_user(u64_to_user_ptr(cmd.response),
 871			 &resp, sizeof(resp)))
 
 872		ret = -EFAULT;
 873
 874	ucma_put_ctx(ctx);
 875	return ret;
 876}
 877
 878static void ucma_query_device_addr(struct rdma_cm_id *cm_id,
 879				   struct rdma_ucm_query_addr_resp *resp)
 880{
 881	if (!cm_id->device)
 882		return;
 883
 884	resp->node_guid = (__force __u64) cm_id->device->node_guid;
 
 885	resp->port_num = cm_id->port_num;
 886	resp->pkey = (__force __u16) cpu_to_be16(
 887		     ib_addr_get_pkey(&cm_id->route.addr.dev_addr));
 888}
 889
 890static ssize_t ucma_query_addr(struct ucma_context *ctx,
 891			       void __user *response, int out_len)
 892{
 893	struct rdma_ucm_query_addr_resp resp;
 894	struct sockaddr *addr;
 895	int ret = 0;
 896
 897	if (out_len < sizeof(resp))
 898		return -ENOSPC;
 899
 900	memset(&resp, 0, sizeof resp);
 901
 902	addr = (struct sockaddr *) &ctx->cm_id->route.addr.src_addr;
 903	resp.src_size = rdma_addr_size(addr);
 904	memcpy(&resp.src_addr, addr, resp.src_size);
 905
 906	addr = (struct sockaddr *) &ctx->cm_id->route.addr.dst_addr;
 907	resp.dst_size = rdma_addr_size(addr);
 908	memcpy(&resp.dst_addr, addr, resp.dst_size);
 909
 910	ucma_query_device_addr(ctx->cm_id, &resp);
 911
 912	if (copy_to_user(response, &resp, sizeof(resp)))
 913		ret = -EFAULT;
 914
 915	return ret;
 916}
 917
 918static ssize_t ucma_query_path(struct ucma_context *ctx,
 919			       void __user *response, int out_len)
 920{
 921	struct rdma_ucm_query_path_resp *resp;
 922	int i, ret = 0;
 923
 924	if (out_len < sizeof(*resp))
 925		return -ENOSPC;
 926
 927	resp = kzalloc(out_len, GFP_KERNEL);
 928	if (!resp)
 929		return -ENOMEM;
 930
 931	resp->num_paths = ctx->cm_id->route.num_paths;
 932	for (i = 0, out_len -= sizeof(*resp);
 933	     i < resp->num_paths && out_len > sizeof(struct ib_path_rec_data);
 934	     i++, out_len -= sizeof(struct ib_path_rec_data)) {
 935		struct sa_path_rec *rec = &ctx->cm_id->route.path_rec[i];
 936
 937		resp->path_data[i].flags = IB_PATH_GMP | IB_PATH_PRIMARY |
 938					   IB_PATH_BIDIRECTIONAL;
 939		if (rec->rec_type == SA_PATH_REC_TYPE_OPA) {
 940			struct sa_path_rec ib;
 941
 942			sa_convert_path_opa_to_ib(&ib, rec);
 943			ib_sa_pack_path(&ib, &resp->path_data[i].path_rec);
 944
 945		} else {
 946			ib_sa_pack_path(rec, &resp->path_data[i].path_rec);
 947		}
 948	}
 949
 950	if (copy_to_user(response, resp,
 951			 sizeof(*resp) + (i * sizeof(struct ib_path_rec_data))))
 952		ret = -EFAULT;
 953
 954	kfree(resp);
 955	return ret;
 956}
 957
 958static ssize_t ucma_query_gid(struct ucma_context *ctx,
 959			      void __user *response, int out_len)
 960{
 961	struct rdma_ucm_query_addr_resp resp;
 962	struct sockaddr_ib *addr;
 963	int ret = 0;
 964
 965	if (out_len < sizeof(resp))
 966		return -ENOSPC;
 967
 968	memset(&resp, 0, sizeof resp);
 969
 970	ucma_query_device_addr(ctx->cm_id, &resp);
 971
 972	addr = (struct sockaddr_ib *) &resp.src_addr;
 973	resp.src_size = sizeof(*addr);
 974	if (ctx->cm_id->route.addr.src_addr.ss_family == AF_IB) {
 975		memcpy(addr, &ctx->cm_id->route.addr.src_addr, resp.src_size);
 976	} else {
 977		addr->sib_family = AF_IB;
 978		addr->sib_pkey = (__force __be16) resp.pkey;
 979		rdma_read_gids(ctx->cm_id, (union ib_gid *)&addr->sib_addr,
 980			       NULL);
 981		addr->sib_sid = rdma_get_service_id(ctx->cm_id, (struct sockaddr *)
 982						    &ctx->cm_id->route.addr.src_addr);
 983	}
 984
 985	addr = (struct sockaddr_ib *) &resp.dst_addr;
 986	resp.dst_size = sizeof(*addr);
 987	if (ctx->cm_id->route.addr.dst_addr.ss_family == AF_IB) {
 988		memcpy(addr, &ctx->cm_id->route.addr.dst_addr, resp.dst_size);
 989	} else {
 990		addr->sib_family = AF_IB;
 991		addr->sib_pkey = (__force __be16) resp.pkey;
 992		rdma_read_gids(ctx->cm_id, NULL,
 993			       (union ib_gid *)&addr->sib_addr);
 994		addr->sib_sid = rdma_get_service_id(ctx->cm_id, (struct sockaddr *)
 995						    &ctx->cm_id->route.addr.dst_addr);
 996	}
 997
 998	if (copy_to_user(response, &resp, sizeof(resp)))
 999		ret = -EFAULT;
1000
1001	return ret;
1002}
1003
1004static ssize_t ucma_query(struct ucma_file *file,
1005			  const char __user *inbuf,
1006			  int in_len, int out_len)
1007{
1008	struct rdma_ucm_query cmd;
1009	struct ucma_context *ctx;
1010	void __user *response;
1011	int ret;
1012
1013	if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
1014		return -EFAULT;
1015
1016	response = u64_to_user_ptr(cmd.response);
1017	ctx = ucma_get_ctx(file, cmd.id);
1018	if (IS_ERR(ctx))
1019		return PTR_ERR(ctx);
1020
 
1021	switch (cmd.option) {
1022	case RDMA_USER_CM_QUERY_ADDR:
1023		ret = ucma_query_addr(ctx, response, out_len);
1024		break;
1025	case RDMA_USER_CM_QUERY_PATH:
1026		ret = ucma_query_path(ctx, response, out_len);
1027		break;
1028	case RDMA_USER_CM_QUERY_GID:
1029		ret = ucma_query_gid(ctx, response, out_len);
1030		break;
1031	default:
1032		ret = -ENOSYS;
1033		break;
1034	}
 
1035
1036	ucma_put_ctx(ctx);
1037	return ret;
1038}
1039
1040static void ucma_copy_conn_param(struct rdma_cm_id *id,
1041				 struct rdma_conn_param *dst,
1042				 struct rdma_ucm_conn_param *src)
1043{
1044	dst->private_data = src->private_data;
1045	dst->private_data_len = src->private_data_len;
1046	dst->responder_resources =src->responder_resources;
1047	dst->initiator_depth = src->initiator_depth;
1048	dst->flow_control = src->flow_control;
1049	dst->retry_count = src->retry_count;
1050	dst->rnr_retry_count = src->rnr_retry_count;
1051	dst->srq = src->srq;
1052	dst->qp_num = src->qp_num;
1053	dst->qkey = (id->route.addr.src_addr.ss_family == AF_IB) ? src->qkey : 0;
1054}
1055
1056static ssize_t ucma_connect(struct ucma_file *file, const char __user *inbuf,
1057			    int in_len, int out_len)
1058{
 
 
1059	struct rdma_ucm_connect cmd;
1060	struct rdma_conn_param conn_param;
1061	struct ucma_context *ctx;
 
1062	int ret;
1063
1064	if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
 
 
 
1065		return -EFAULT;
1066
1067	if (!cmd.conn_param.valid)
1068		return -EINVAL;
1069
1070	ctx = ucma_get_ctx_dev(file, cmd.id);
1071	if (IS_ERR(ctx))
1072		return PTR_ERR(ctx);
1073
1074	ucma_copy_conn_param(ctx->cm_id, &conn_param, &cmd.conn_param);
1075	ret = rdma_connect(ctx->cm_id, &conn_param);
 
 
 
 
 
 
 
1076	ucma_put_ctx(ctx);
1077	return ret;
1078}
1079
1080static ssize_t ucma_listen(struct ucma_file *file, const char __user *inbuf,
1081			   int in_len, int out_len)
1082{
1083	struct rdma_ucm_listen cmd;
1084	struct ucma_context *ctx;
1085	int ret;
1086
1087	if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
1088		return -EFAULT;
1089
1090	ctx = ucma_get_ctx(file, cmd.id);
1091	if (IS_ERR(ctx))
1092		return PTR_ERR(ctx);
1093
1094	ctx->backlog = cmd.backlog > 0 && cmd.backlog < max_backlog ?
1095		       cmd.backlog : max_backlog;
1096	ret = rdma_listen(ctx->cm_id, ctx->backlog);
 
 
 
 
1097	ucma_put_ctx(ctx);
1098	return ret;
1099}
1100
1101static ssize_t ucma_accept(struct ucma_file *file, const char __user *inbuf,
1102			   int in_len, int out_len)
1103{
1104	struct rdma_ucm_accept cmd;
1105	struct rdma_conn_param conn_param;
 
1106	struct ucma_context *ctx;
 
1107	int ret;
1108
1109	if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
 
 
 
1110		return -EFAULT;
1111
1112	ctx = ucma_get_ctx_dev(file, cmd.id);
1113	if (IS_ERR(ctx))
1114		return PTR_ERR(ctx);
1115
 
 
 
 
 
1116	if (cmd.conn_param.valid) {
1117		ucma_copy_conn_param(ctx->cm_id, &conn_param, &cmd.conn_param);
1118		mutex_lock(&file->mut);
1119		ret = __rdma_accept(ctx->cm_id, &conn_param, NULL);
1120		if (!ret)
 
 
1121			ctx->uid = cmd.uid;
1122		mutex_unlock(&file->mut);
1123	} else
1124		ret = __rdma_accept(ctx->cm_id, NULL, NULL);
1125
 
 
 
 
 
 
1126	ucma_put_ctx(ctx);
1127	return ret;
1128}
1129
1130static ssize_t ucma_reject(struct ucma_file *file, const char __user *inbuf,
1131			   int in_len, int out_len)
1132{
1133	struct rdma_ucm_reject cmd;
1134	struct ucma_context *ctx;
1135	int ret;
1136
1137	if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
1138		return -EFAULT;
1139
 
 
 
 
 
 
 
 
 
 
 
1140	ctx = ucma_get_ctx_dev(file, cmd.id);
1141	if (IS_ERR(ctx))
1142		return PTR_ERR(ctx);
1143
1144	ret = rdma_reject(ctx->cm_id, cmd.private_data, cmd.private_data_len);
 
 
 
1145	ucma_put_ctx(ctx);
1146	return ret;
1147}
1148
1149static ssize_t ucma_disconnect(struct ucma_file *file, const char __user *inbuf,
1150			       int in_len, int out_len)
1151{
1152	struct rdma_ucm_disconnect cmd;
1153	struct ucma_context *ctx;
1154	int ret;
1155
1156	if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
1157		return -EFAULT;
1158
1159	ctx = ucma_get_ctx_dev(file, cmd.id);
1160	if (IS_ERR(ctx))
1161		return PTR_ERR(ctx);
1162
 
1163	ret = rdma_disconnect(ctx->cm_id);
 
1164	ucma_put_ctx(ctx);
1165	return ret;
1166}
1167
1168static ssize_t ucma_init_qp_attr(struct ucma_file *file,
1169				 const char __user *inbuf,
1170				 int in_len, int out_len)
1171{
1172	struct rdma_ucm_init_qp_attr cmd;
1173	struct ib_uverbs_qp_attr resp;
1174	struct ucma_context *ctx;
1175	struct ib_qp_attr qp_attr;
1176	int ret;
1177
1178	if (out_len < sizeof(resp))
1179		return -ENOSPC;
1180
1181	if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
1182		return -EFAULT;
1183
1184	if (cmd.qp_state > IB_QPS_ERR)
1185		return -EINVAL;
1186
1187	ctx = ucma_get_ctx_dev(file, cmd.id);
1188	if (IS_ERR(ctx))
1189		return PTR_ERR(ctx);
1190
1191	resp.qp_attr_mask = 0;
1192	memset(&qp_attr, 0, sizeof qp_attr);
1193	qp_attr.qp_state = cmd.qp_state;
 
1194	ret = rdma_init_qp_attr(ctx->cm_id, &qp_attr, &resp.qp_attr_mask);
 
1195	if (ret)
1196		goto out;
1197
1198	ib_copy_qp_attr_to_user(ctx->cm_id->device, &resp, &qp_attr);
1199	if (copy_to_user(u64_to_user_ptr(cmd.response),
1200			 &resp, sizeof(resp)))
1201		ret = -EFAULT;
1202
1203out:
1204	ucma_put_ctx(ctx);
1205	return ret;
1206}
1207
1208static int ucma_set_option_id(struct ucma_context *ctx, int optname,
1209			      void *optval, size_t optlen)
1210{
1211	int ret = 0;
1212
1213	switch (optname) {
1214	case RDMA_OPTION_ID_TOS:
1215		if (optlen != sizeof(u8)) {
1216			ret = -EINVAL;
1217			break;
1218		}
1219		rdma_set_service_type(ctx->cm_id, *((u8 *) optval));
1220		break;
1221	case RDMA_OPTION_ID_REUSEADDR:
1222		if (optlen != sizeof(int)) {
1223			ret = -EINVAL;
1224			break;
1225		}
1226		ret = rdma_set_reuseaddr(ctx->cm_id, *((int *) optval) ? 1 : 0);
1227		break;
1228	case RDMA_OPTION_ID_AFONLY:
1229		if (optlen != sizeof(int)) {
1230			ret = -EINVAL;
1231			break;
1232		}
1233		ret = rdma_set_afonly(ctx->cm_id, *((int *) optval) ? 1 : 0);
1234		break;
 
 
 
 
 
 
 
1235	default:
1236		ret = -ENOSYS;
1237	}
1238
1239	return ret;
1240}
1241
1242static int ucma_set_ib_path(struct ucma_context *ctx,
1243			    struct ib_path_rec_data *path_data, size_t optlen)
1244{
1245	struct sa_path_rec sa_path;
1246	struct rdma_cm_event event;
1247	int ret;
1248
1249	if (optlen % sizeof(*path_data))
1250		return -EINVAL;
1251
1252	for (; optlen; optlen -= sizeof(*path_data), path_data++) {
1253		if (path_data->flags == (IB_PATH_GMP | IB_PATH_PRIMARY |
1254					 IB_PATH_BIDIRECTIONAL))
1255			break;
1256	}
1257
1258	if (!optlen)
1259		return -EINVAL;
1260
1261	if (!ctx->cm_id->device)
1262		return -EINVAL;
1263
1264	memset(&sa_path, 0, sizeof(sa_path));
1265
1266	sa_path.rec_type = SA_PATH_REC_TYPE_IB;
1267	ib_sa_unpack_path(path_data->path_rec, &sa_path);
1268
1269	if (rdma_cap_opa_ah(ctx->cm_id->device, ctx->cm_id->port_num)) {
1270		struct sa_path_rec opa;
1271
1272		sa_convert_path_ib_to_opa(&opa, &sa_path);
 
1273		ret = rdma_set_ib_path(ctx->cm_id, &opa);
 
1274	} else {
 
1275		ret = rdma_set_ib_path(ctx->cm_id, &sa_path);
 
1276	}
1277	if (ret)
1278		return ret;
1279
1280	memset(&event, 0, sizeof event);
1281	event.event = RDMA_CM_EVENT_ROUTE_RESOLVED;
1282	return ucma_event_handler(ctx->cm_id, &event);
1283}
1284
1285static int ucma_set_option_ib(struct ucma_context *ctx, int optname,
1286			      void *optval, size_t optlen)
1287{
1288	int ret;
1289
1290	switch (optname) {
1291	case RDMA_OPTION_IB_PATH:
1292		ret = ucma_set_ib_path(ctx, optval, optlen);
1293		break;
1294	default:
1295		ret = -ENOSYS;
1296	}
1297
1298	return ret;
1299}
1300
1301static int ucma_set_option_level(struct ucma_context *ctx, int level,
1302				 int optname, void *optval, size_t optlen)
1303{
1304	int ret;
1305
1306	switch (level) {
1307	case RDMA_OPTION_ID:
 
1308		ret = ucma_set_option_id(ctx, optname, optval, optlen);
 
1309		break;
1310	case RDMA_OPTION_IB:
1311		ret = ucma_set_option_ib(ctx, optname, optval, optlen);
1312		break;
1313	default:
1314		ret = -ENOSYS;
1315	}
1316
1317	return ret;
1318}
1319
1320static ssize_t ucma_set_option(struct ucma_file *file, const char __user *inbuf,
1321			       int in_len, int out_len)
1322{
1323	struct rdma_ucm_set_option cmd;
1324	struct ucma_context *ctx;
1325	void *optval;
1326	int ret;
1327
1328	if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
1329		return -EFAULT;
1330
1331	if (unlikely(cmd.optlen > KMALLOC_MAX_SIZE))
1332		return -EINVAL;
1333
1334	ctx = ucma_get_ctx(file, cmd.id);
1335	if (IS_ERR(ctx))
1336		return PTR_ERR(ctx);
1337
1338	optval = memdup_user(u64_to_user_ptr(cmd.optval),
1339			     cmd.optlen);
1340	if (IS_ERR(optval)) {
1341		ret = PTR_ERR(optval);
1342		goto out;
1343	}
1344
1345	ret = ucma_set_option_level(ctx, cmd.level, cmd.optname, optval,
1346				    cmd.optlen);
1347	kfree(optval);
1348
1349out:
1350	ucma_put_ctx(ctx);
1351	return ret;
1352}
1353
1354static ssize_t ucma_notify(struct ucma_file *file, const char __user *inbuf,
1355			   int in_len, int out_len)
1356{
1357	struct rdma_ucm_notify cmd;
1358	struct ucma_context *ctx;
1359	int ret = -EINVAL;
1360
1361	if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
1362		return -EFAULT;
1363
1364	ctx = ucma_get_ctx(file, cmd.id);
1365	if (IS_ERR(ctx))
1366		return PTR_ERR(ctx);
1367
 
1368	if (ctx->cm_id->device)
1369		ret = rdma_notify(ctx->cm_id, (enum ib_event_type)cmd.event);
 
1370
1371	ucma_put_ctx(ctx);
1372	return ret;
1373}
1374
1375static ssize_t ucma_process_join(struct ucma_file *file,
1376				 struct rdma_ucm_join_mcast *cmd,  int out_len)
1377{
1378	struct rdma_ucm_create_id_resp resp;
1379	struct ucma_context *ctx;
1380	struct ucma_multicast *mc;
1381	struct sockaddr *addr;
1382	int ret;
1383	u8 join_state;
1384
1385	if (out_len < sizeof(resp))
1386		return -ENOSPC;
1387
1388	addr = (struct sockaddr *) &cmd->addr;
1389	if (cmd->addr_size != rdma_addr_size(addr))
1390		return -EINVAL;
1391
1392	if (cmd->join_flags == RDMA_MC_JOIN_FLAG_FULLMEMBER)
1393		join_state = BIT(FULLMEMBER_JOIN);
1394	else if (cmd->join_flags == RDMA_MC_JOIN_FLAG_SENDONLY_FULLMEMBER)
1395		join_state = BIT(SENDONLY_FULLMEMBER_JOIN);
1396	else
1397		return -EINVAL;
1398
1399	ctx = ucma_get_ctx_dev(file, cmd->id);
1400	if (IS_ERR(ctx))
1401		return PTR_ERR(ctx);
1402
1403	mutex_lock(&file->mut);
1404	mc = ucma_alloc_multicast(ctx);
1405	if (!mc) {
1406		ret = -ENOMEM;
1407		goto err1;
1408	}
 
 
1409	mc->join_state = join_state;
1410	mc->uid = cmd->uid;
1411	memcpy(&mc->addr, addr, cmd->addr_size);
 
 
 
 
 
 
 
 
 
 
 
 
1412	ret = rdma_join_multicast(ctx->cm_id, (struct sockaddr *)&mc->addr,
1413				  join_state, mc);
 
1414	if (ret)
1415		goto err2;
1416
1417	resp.id = mc->id;
1418	if (copy_to_user(u64_to_user_ptr(cmd->response),
1419			 &resp, sizeof(resp))) {
1420		ret = -EFAULT;
1421		goto err3;
1422	}
1423
1424	mutex_unlock(&file->mut);
 
1425	ucma_put_ctx(ctx);
1426	return 0;
1427
1428err3:
 
1429	rdma_leave_multicast(ctx->cm_id, (struct sockaddr *) &mc->addr);
 
1430	ucma_cleanup_mc_events(mc);
1431err2:
1432	mutex_lock(&mut);
1433	idr_remove(&multicast_idr, mc->id);
1434	mutex_unlock(&mut);
1435	list_del(&mc->list);
 
 
 
1436	kfree(mc);
1437err1:
1438	mutex_unlock(&file->mut);
1439	ucma_put_ctx(ctx);
1440	return ret;
1441}
1442
1443static ssize_t ucma_join_ip_multicast(struct ucma_file *file,
1444				      const char __user *inbuf,
1445				      int in_len, int out_len)
1446{
1447	struct rdma_ucm_join_ip_mcast cmd;
1448	struct rdma_ucm_join_mcast join_cmd;
1449
1450	if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
1451		return -EFAULT;
1452
1453	join_cmd.response = cmd.response;
1454	join_cmd.uid = cmd.uid;
1455	join_cmd.id = cmd.id;
1456	join_cmd.addr_size = rdma_addr_size_in6(&cmd.addr);
1457	if (!join_cmd.addr_size)
1458		return -EINVAL;
1459
1460	join_cmd.join_flags = RDMA_MC_JOIN_FLAG_FULLMEMBER;
1461	memcpy(&join_cmd.addr, &cmd.addr, join_cmd.addr_size);
1462
1463	return ucma_process_join(file, &join_cmd, out_len);
1464}
1465
1466static ssize_t ucma_join_multicast(struct ucma_file *file,
1467				   const char __user *inbuf,
1468				   int in_len, int out_len)
1469{
1470	struct rdma_ucm_join_mcast cmd;
1471
1472	if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
1473		return -EFAULT;
1474
1475	if (!rdma_addr_size_kss(&cmd.addr))
1476		return -EINVAL;
1477
1478	return ucma_process_join(file, &cmd, out_len);
1479}
1480
1481static ssize_t ucma_leave_multicast(struct ucma_file *file,
1482				    const char __user *inbuf,
1483				    int in_len, int out_len)
1484{
1485	struct rdma_ucm_destroy_id cmd;
1486	struct rdma_ucm_destroy_id_resp resp;
1487	struct ucma_multicast *mc;
1488	int ret = 0;
1489
1490	if (out_len < sizeof(resp))
1491		return -ENOSPC;
1492
1493	if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
1494		return -EFAULT;
1495
1496	mutex_lock(&mut);
1497	mc = idr_find(&multicast_idr, cmd.id);
1498	if (!mc)
1499		mc = ERR_PTR(-ENOENT);
1500	else if (mc->ctx->file != file)
1501		mc = ERR_PTR(-EINVAL);
1502	else if (!atomic_inc_not_zero(&mc->ctx->ref))
1503		mc = ERR_PTR(-ENXIO);
1504	else
1505		idr_remove(&multicast_idr, mc->id);
1506	mutex_unlock(&mut);
1507
1508	if (IS_ERR(mc)) {
 
1509		ret = PTR_ERR(mc);
1510		goto out;
1511	}
1512
 
 
 
 
 
1513	rdma_leave_multicast(mc->ctx->cm_id, (struct sockaddr *) &mc->addr);
1514	mutex_lock(&mc->ctx->file->mut);
 
1515	ucma_cleanup_mc_events(mc);
1516	list_del(&mc->list);
1517	mutex_unlock(&mc->ctx->file->mut);
1518
1519	ucma_put_ctx(mc->ctx);
1520	resp.events_reported = mc->events_reported;
1521	kfree(mc);
1522
1523	if (copy_to_user(u64_to_user_ptr(cmd.response),
1524			 &resp, sizeof(resp)))
1525		ret = -EFAULT;
1526out:
1527	return ret;
1528}
1529
1530static void ucma_lock_files(struct ucma_file *file1, struct ucma_file *file2)
1531{
1532	/* Acquire mutex's based on pointer comparison to prevent deadlock. */
1533	if (file1 < file2) {
1534		mutex_lock(&file1->mut);
1535		mutex_lock_nested(&file2->mut, SINGLE_DEPTH_NESTING);
1536	} else {
1537		mutex_lock(&file2->mut);
1538		mutex_lock_nested(&file1->mut, SINGLE_DEPTH_NESTING);
1539	}
1540}
1541
1542static void ucma_unlock_files(struct ucma_file *file1, struct ucma_file *file2)
1543{
1544	if (file1 < file2) {
1545		mutex_unlock(&file2->mut);
1546		mutex_unlock(&file1->mut);
1547	} else {
1548		mutex_unlock(&file1->mut);
1549		mutex_unlock(&file2->mut);
1550	}
1551}
1552
1553static void ucma_move_events(struct ucma_context *ctx, struct ucma_file *file)
1554{
1555	struct ucma_event *uevent, *tmp;
1556
1557	list_for_each_entry_safe(uevent, tmp, &ctx->file->event_list, list)
1558		if (uevent->ctx == ctx)
1559			list_move_tail(&uevent->list, &file->event_list);
1560}
1561
1562static ssize_t ucma_migrate_id(struct ucma_file *new_file,
1563			       const char __user *inbuf,
1564			       int in_len, int out_len)
1565{
1566	struct rdma_ucm_migrate_id cmd;
1567	struct rdma_ucm_migrate_resp resp;
 
1568	struct ucma_context *ctx;
 
1569	struct fd f;
1570	struct ucma_file *cur_file;
1571	int ret = 0;
1572
1573	if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
1574		return -EFAULT;
1575
1576	/* Get current fd to protect against it being closed */
1577	f = fdget(cmd.fd);
1578	if (!f.file)
1579		return -ENOENT;
 
 
 
 
 
1580
1581	/* Validate current fd and prevent destruction of id. */
1582	ctx = ucma_get_ctx(f.file->private_data, cmd.id);
1583	if (IS_ERR(ctx)) {
1584		ret = PTR_ERR(ctx);
1585		goto file_put;
1586	}
1587
1588	cur_file = ctx->file;
1589	if (cur_file == new_file) {
1590		resp.events_reported = ctx->events_reported;
1591		goto response;
 
 
 
 
 
 
 
1592	}
 
 
1593
 
 
1594	/*
1595	 * Migrate events between fd's, maintaining order, and avoiding new
1596	 * events being added before existing events.
1597	 */
1598	ucma_lock_files(cur_file, new_file);
1599	mutex_lock(&mut);
1600
1601	list_move_tail(&ctx->list, &new_file->ctx_list);
1602	ucma_move_events(ctx, new_file);
1603	ctx->file = new_file;
1604	resp.events_reported = ctx->events_reported;
 
1605
1606	mutex_unlock(&mut);
1607	ucma_unlock_files(cur_file, new_file);
 
 
1608
1609response:
1610	if (copy_to_user(u64_to_user_ptr(cmd.response),
1611			 &resp, sizeof(resp)))
1612		ret = -EFAULT;
1613
 
 
1614	ucma_put_ctx(ctx);
1615file_put:
1616	fdput(f);
1617	return ret;
1618}
1619
1620static ssize_t (*ucma_cmd_table[])(struct ucma_file *file,
1621				   const char __user *inbuf,
1622				   int in_len, int out_len) = {
1623	[RDMA_USER_CM_CMD_CREATE_ID] 	 = ucma_create_id,
1624	[RDMA_USER_CM_CMD_DESTROY_ID]	 = ucma_destroy_id,
1625	[RDMA_USER_CM_CMD_BIND_IP]	 = ucma_bind_ip,
1626	[RDMA_USER_CM_CMD_RESOLVE_IP]	 = ucma_resolve_ip,
1627	[RDMA_USER_CM_CMD_RESOLVE_ROUTE] = ucma_resolve_route,
1628	[RDMA_USER_CM_CMD_QUERY_ROUTE]	 = ucma_query_route,
1629	[RDMA_USER_CM_CMD_CONNECT]	 = ucma_connect,
1630	[RDMA_USER_CM_CMD_LISTEN]	 = ucma_listen,
1631	[RDMA_USER_CM_CMD_ACCEPT]	 = ucma_accept,
1632	[RDMA_USER_CM_CMD_REJECT]	 = ucma_reject,
1633	[RDMA_USER_CM_CMD_DISCONNECT]	 = ucma_disconnect,
1634	[RDMA_USER_CM_CMD_INIT_QP_ATTR]	 = ucma_init_qp_attr,
1635	[RDMA_USER_CM_CMD_GET_EVENT]	 = ucma_get_event,
1636	[RDMA_USER_CM_CMD_GET_OPTION]	 = NULL,
1637	[RDMA_USER_CM_CMD_SET_OPTION]	 = ucma_set_option,
1638	[RDMA_USER_CM_CMD_NOTIFY]	 = ucma_notify,
1639	[RDMA_USER_CM_CMD_JOIN_IP_MCAST] = ucma_join_ip_multicast,
1640	[RDMA_USER_CM_CMD_LEAVE_MCAST]	 = ucma_leave_multicast,
1641	[RDMA_USER_CM_CMD_MIGRATE_ID]	 = ucma_migrate_id,
1642	[RDMA_USER_CM_CMD_QUERY]	 = ucma_query,
1643	[RDMA_USER_CM_CMD_BIND]		 = ucma_bind,
1644	[RDMA_USER_CM_CMD_RESOLVE_ADDR]	 = ucma_resolve_addr,
1645	[RDMA_USER_CM_CMD_JOIN_MCAST]	 = ucma_join_multicast
1646};
1647
1648static ssize_t ucma_write(struct file *filp, const char __user *buf,
1649			  size_t len, loff_t *pos)
1650{
1651	struct ucma_file *file = filp->private_data;
1652	struct rdma_ucm_cmd_hdr hdr;
1653	ssize_t ret;
1654
1655	if (!ib_safe_file_access(filp)) {
1656		pr_err_once("ucma_write: process %d (%s) changed security contexts after opening file descriptor, this is not allowed.\n",
1657			    task_tgid_vnr(current), current->comm);
1658		return -EACCES;
1659	}
1660
1661	if (len < sizeof(hdr))
1662		return -EINVAL;
1663
1664	if (copy_from_user(&hdr, buf, sizeof(hdr)))
1665		return -EFAULT;
1666
1667	if (hdr.cmd >= ARRAY_SIZE(ucma_cmd_table))
1668		return -EINVAL;
 
1669
1670	if (hdr.in + sizeof(hdr) > len)
1671		return -EINVAL;
1672
1673	if (!ucma_cmd_table[hdr.cmd])
1674		return -ENOSYS;
1675
1676	ret = ucma_cmd_table[hdr.cmd](file, buf + sizeof(hdr), hdr.in, hdr.out);
1677	if (!ret)
1678		ret = len;
1679
1680	return ret;
1681}
1682
1683static __poll_t ucma_poll(struct file *filp, struct poll_table_struct *wait)
1684{
1685	struct ucma_file *file = filp->private_data;
1686	__poll_t mask = 0;
1687
1688	poll_wait(filp, &file->poll_wait, wait);
1689
1690	if (!list_empty(&file->event_list))
1691		mask = EPOLLIN | EPOLLRDNORM;
1692
1693	return mask;
1694}
1695
1696/*
1697 * ucma_open() does not need the BKL:
1698 *
1699 *  - no global state is referred to;
1700 *  - there is no ioctl method to race against;
1701 *  - no further module initialization is required for open to work
1702 *    after the device is registered.
1703 */
1704static int ucma_open(struct inode *inode, struct file *filp)
1705{
1706	struct ucma_file *file;
1707
1708	file = kmalloc(sizeof *file, GFP_KERNEL);
1709	if (!file)
1710		return -ENOMEM;
1711
1712	file->close_wq = alloc_ordered_workqueue("ucma_close_id",
1713						 WQ_MEM_RECLAIM);
1714	if (!file->close_wq) {
1715		kfree(file);
1716		return -ENOMEM;
1717	}
1718
1719	INIT_LIST_HEAD(&file->event_list);
1720	INIT_LIST_HEAD(&file->ctx_list);
1721	init_waitqueue_head(&file->poll_wait);
1722	mutex_init(&file->mut);
1723
1724	filp->private_data = file;
1725	file->filp = filp;
1726
1727	return nonseekable_open(inode, filp);
1728}
1729
1730static int ucma_close(struct inode *inode, struct file *filp)
1731{
1732	struct ucma_file *file = filp->private_data;
1733	struct ucma_context *ctx, *tmp;
1734
1735	mutex_lock(&file->mut);
1736	list_for_each_entry_safe(ctx, tmp, &file->ctx_list, list) {
1737		ctx->destroying = 1;
1738		mutex_unlock(&file->mut);
1739
1740		mutex_lock(&mut);
1741		idr_remove(&ctx_idr, ctx->id);
1742		mutex_unlock(&mut);
1743
1744		flush_workqueue(file->close_wq);
1745		/* At that step once ctx was marked as destroying and workqueue
1746		 * was flushed we are safe from any inflights handlers that
1747		 * might put other closing task.
1748		 */
1749		mutex_lock(&mut);
1750		if (!ctx->closing) {
1751			mutex_unlock(&mut);
1752			/* rdma_destroy_id ensures that no event handlers are
1753			 * inflight for that id before releasing it.
1754			 */
1755			rdma_destroy_id(ctx->cm_id);
1756		} else {
1757			mutex_unlock(&mut);
1758		}
1759
1760		ucma_free_ctx(ctx);
1761		mutex_lock(&file->mut);
1762	}
1763	mutex_unlock(&file->mut);
1764	destroy_workqueue(file->close_wq);
1765	kfree(file);
1766	return 0;
1767}
1768
1769static const struct file_operations ucma_fops = {
1770	.owner 	 = THIS_MODULE,
1771	.open 	 = ucma_open,
1772	.release = ucma_close,
1773	.write	 = ucma_write,
1774	.poll    = ucma_poll,
1775	.llseek	 = no_llseek,
1776};
1777
1778static struct miscdevice ucma_misc = {
1779	.minor		= MISC_DYNAMIC_MINOR,
1780	.name		= "rdma_cm",
1781	.nodename	= "infiniband/rdma_cm",
1782	.mode		= 0666,
1783	.fops		= &ucma_fops,
1784};
1785
1786static ssize_t show_abi_version(struct device *dev,
1787				struct device_attribute *attr,
1788				char *buf)
 
 
 
 
 
 
 
 
 
 
 
 
1789{
1790	return sprintf(buf, "%d\n", RDMA_USER_CM_ABI_VERSION);
1791}
1792static DEVICE_ATTR(abi_version, S_IRUGO, show_abi_version, NULL);
1793
1794static int __init ucma_init(void)
1795{
1796	int ret;
1797
1798	ret = misc_register(&ucma_misc);
1799	if (ret)
1800		return ret;
1801
1802	ret = device_create_file(ucma_misc.this_device, &dev_attr_abi_version);
1803	if (ret) {
1804		pr_err("rdma_ucm: couldn't create abi_version attr\n");
1805		goto err1;
1806	}
1807
1808	ucma_ctl_table_hdr = register_net_sysctl(&init_net, "net/rdma_ucm", ucma_ctl_table);
1809	if (!ucma_ctl_table_hdr) {
1810		pr_err("rdma_ucm: couldn't register sysctl paths\n");
1811		ret = -ENOMEM;
1812		goto err2;
1813	}
 
 
 
 
 
1814	return 0;
 
 
1815err2:
1816	device_remove_file(ucma_misc.this_device, &dev_attr_abi_version);
1817err1:
1818	misc_deregister(&ucma_misc);
1819	return ret;
1820}
1821
1822static void __exit ucma_cleanup(void)
1823{
 
1824	unregister_net_sysctl_table(ucma_ctl_table_hdr);
1825	device_remove_file(ucma_misc.this_device, &dev_attr_abi_version);
1826	misc_deregister(&ucma_misc);
1827	idr_destroy(&ctx_idr);
1828	idr_destroy(&multicast_idr);
1829}
1830
1831module_init(ucma_init);
1832module_exit(ucma_cleanup);