brintos

brintos / linux-shallow public Read only

0
0
Text · 46.9 KiB · 5dbb248 Raw
1895 lines · c
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 two5 * licenses.  You may choose to be licensed under the terms of the GNU6 * General Public License (GPL) Version 2, available from the file7 * COPYING in the main directory of this source tree, or the8 * OpenIB.org BSD license below:9 *10 *     Redistribution and use in source and binary forms, with or11 *     without modification, are permitted provided that the following12 *     conditions are met:13 *14 *      - Redistributions of source code must retain the above15 *	copyright notice, this list of conditions and the following16 *	disclaimer.17 *18 *      - Redistributions in binary form must reproduce the above19 *	copyright notice, this list of conditions and the following20 *	disclaimer in the documentation and/or other materials21 *	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 OF25 * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND26 * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS27 * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN28 * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN29 * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE30 * 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 the160 * 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 underlying180	 * resources. The context is still alive till its explicit destryoing181	 * 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	else287		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 their343	 * context.  This can only happen if an error occurs on a new connection344	 * before the user accepts it.  This is okay, since the accept will just345	 * fail later. However, we do need to release the underlying HW346	 * 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 the376	 * 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 the496		 * lock on the reader and this is enough serialization497		 */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 that549	 * have not been delivered to userspace are cleaned up too. Must be done550	 * 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 (ie561 * the ctx is not public to the user). This either because:562 *  - ucma_finish_ctx() hasn't been called563 *  - 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 by571	 * the removal from the xarray. Once the work is cancled ref will either572	 * be 0 because the work ran to completion and consumed the ref from the573	 * xarray, or it will be positive because we still have the ref from the574	 * xarray. This can also be 0 in cases where cm_id was never set575	 */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	else1457		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 (!fd_file(f))1628		return -ENOENT;1629	if (fd_file(f)->f_op != &ucma_fops) {1630		ret = -EINVAL;1631		goto file_put;1632	}1633	cur_file = fd_file(f)->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 destruction1646	 * 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 for1661	 * 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_multicast1712};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 work1769 *    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() are1796	 * prevented by this being a FD release function. The list_add_tail() in1797	 * ucma_connect_event_handler() can run concurrently, however it only1798	 * adds to the list *after* a listening ID. By only reading the first of1799	 * the list, and relying on ucma_destroy_private_ctx() to block1800	 * 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};1821 1822static struct miscdevice ucma_misc = {1823	.minor		= MISC_DYNAMIC_MINOR,1824	.name		= "rdma_cm",1825	.nodename	= "infiniband/rdma_cm",1826	.mode		= 0666,1827	.fops		= &ucma_fops,1828};1829 1830static int ucma_get_global_nl_info(struct ib_client_nl_info *res)1831{1832	res->abi = RDMA_USER_CM_ABI_VERSION;1833	res->cdev = ucma_misc.this_device;1834	return 0;1835}1836 1837static struct ib_client rdma_cma_client = {1838	.name = "rdma_cm",1839	.get_global_nl_info = ucma_get_global_nl_info,1840};1841MODULE_ALIAS_RDMA_CLIENT("rdma_cm");1842 1843static ssize_t abi_version_show(struct device *dev,1844				struct device_attribute *attr, char *buf)1845{1846	return sysfs_emit(buf, "%d\n", RDMA_USER_CM_ABI_VERSION);1847}1848static DEVICE_ATTR_RO(abi_version);1849 1850static int __init ucma_init(void)1851{1852	int ret;1853 1854	ret = misc_register(&ucma_misc);1855	if (ret)1856		return ret;1857 1858	ret = device_create_file(ucma_misc.this_device, &dev_attr_abi_version);1859	if (ret) {1860		pr_err("rdma_ucm: couldn't create abi_version attr\n");1861		goto err1;1862	}1863 1864	ucma_ctl_table_hdr = register_net_sysctl(&init_net, "net/rdma_ucm", ucma_ctl_table);1865	if (!ucma_ctl_table_hdr) {1866		pr_err("rdma_ucm: couldn't register sysctl paths\n");1867		ret = -ENOMEM;1868		goto err2;1869	}1870 1871	ret = ib_register_client(&rdma_cma_client);1872	if (ret)1873		goto err3;1874 1875	return 0;1876err3:1877	unregister_net_sysctl_table(ucma_ctl_table_hdr);1878err2:1879	device_remove_file(ucma_misc.this_device, &dev_attr_abi_version);1880err1:1881	misc_deregister(&ucma_misc);1882	return ret;1883}1884 1885static void __exit ucma_cleanup(void)1886{1887	ib_unregister_client(&rdma_cma_client);1888	unregister_net_sysctl_table(ucma_ctl_table_hdr);1889	device_remove_file(ucma_misc.this_device, &dev_attr_abi_version);1890	misc_deregister(&ucma_misc);1891}1892 1893module_init(ucma_init);1894module_exit(ucma_cleanup);1895