~hc/RK356X_SDK_RELEASE.git

..	..	@@ -37,73 +37,241 @@
37	37	#include <linux/kernel.h>
38	38	#include <linux/slab.h>
39	39	#include <linux/init.h>
40		-#include <linux/mutex.h>
41	40	#include <linux/netdevice.h>
	41	+#include <net/net_namespace.h>
42	42	#include <linux/security.h>
43	43	#include <linux/notifier.h>
	44	+#include <linux/hashtable.h>
44	45	#include <rdma/rdma_netlink.h>
45	46	#include <rdma/ib_addr.h>
46	47	#include <rdma/ib_cache.h>
	48	+#include <rdma/rdma_counter.h>
47	49
48	50	#include "core_priv.h"
	51	+#include "restrack.h"
49	52
50	53	MODULE_AUTHOR("Roland Dreier");
51	54	MODULE_DESCRIPTION("core kernel InfiniBand API");
52	55	MODULE_LICENSE("Dual BSD/GPL");
53		-
54		-struct ib_client_data {
55		- struct list_head list;
56		- struct ib_client *client;
57		- void * data;
58		- /* The device or client is going down. Do not call client or device
59		- * callbacks other than remove(). */
60		- bool going_down;
61		-};
62	56
63	57	struct workqueue_struct *ib_comp_wq;
64	58	struct workqueue_struct *ib_comp_unbound_wq;
65	59	struct workqueue_struct *ib_wq;
66	60	EXPORT_SYMBOL_GPL(ib_wq);
67	61
68		-/* The device_list and client_list contain devices and clients after their
69		- * registration has completed, and the devices and clients are removed
70		- * during unregistration. */
71		-static LIST_HEAD(device_list);
72		-static LIST_HEAD(client_list);
	62	+/*
	63	+ * Each of the three rwsem locks (devices, clients, client_data) protects the
	64	+ * xarray of the same name. Specifically it allows the caller to assert that
	65	+ * the MARK will/will not be changing under the lock, and for devices and
	66	+ * clients, that the value in the xarray is still a valid pointer. Change of
	67	+ * the MARK is linked to the object state, so holding the lock and testing the
	68	+ * MARK also asserts that the contained object is in a certain state.
	69	+ *
	70	+ * This is used to build a two stage register/unregister flow where objects
	71	+ * can continue to be in the xarray even though they are still in progress to
	72	+ * register/unregister.
	73	+ *
	74	+ * The xarray itself provides additional locking, and restartable iteration,
	75	+ * which is also relied on.
	76	+ *
	77	+ * Locks should not be nested, with the exception of client_data, which is
	78	+ * allowed to nest under the read side of the other two locks.
	79	+ *
	80	+ * The devices_rwsem also protects the device name list, any change or
	81	+ * assignment of device name must also hold the write side to guarantee unique
	82	+ * names.
	83	+ */
73	84
74	85	/*
75		- * device_mutex and lists_rwsem protect access to both device_list and
76		- * client_list. device_mutex protects writer access by device and client
77		- * registration / de-registration. lists_rwsem protects reader access to
78		- * these lists. Iterators of these lists must lock it for read, while updates
79		- * to the lists must be done with a write lock. A special case is when the
80		- * device_mutex is locked. In this case locking the lists for read access is
81		- * not necessary as the device_mutex implies it.
	86	+ * devices contains devices that have had their names assigned. The
	87	+ * devices may not be registered. Users that care about the registration
	88	+ * status need to call ib_device_try_get() on the device to ensure it is
	89	+ * registered, and keep it registered, for the required duration.
82	90	*
83		- * lists_rwsem also protects access to the client data list.
84	91	*/
85		-static DEFINE_MUTEX(device_mutex);
86		-static DECLARE_RWSEM(lists_rwsem);
	92	+static DEFINE_XARRAY_FLAGS(devices, XA_FLAGS_ALLOC);
	93	+static DECLARE_RWSEM(devices_rwsem);
	94	+#define DEVICE_REGISTERED XA_MARK_1
87	95
	96	+static u32 highest_client_id;
	97	+#define CLIENT_REGISTERED XA_MARK_1
	98	+static DEFINE_XARRAY_FLAGS(clients, XA_FLAGS_ALLOC);
	99	+static DECLARE_RWSEM(clients_rwsem);
	100	+
	101	+static void ib_client_put(struct ib_client *client)
	102	+{
	103	+ if (refcount_dec_and_test(&client->uses))
	104	+ complete(&client->uses_zero);
	105	+}
	106	+
	107	+/*
	108	+ * If client_data is registered then the corresponding client must also still
	109	+ * be registered.
	110	+ */
	111	+#define CLIENT_DATA_REGISTERED XA_MARK_1
	112	+
	113	+unsigned int rdma_dev_net_id;
	114	+
	115	+/*
	116	+ * A list of net namespaces is maintained in an xarray. This is necessary
	117	+ * because we can't get the locking right using the existing net ns list. We
	118	+ * would require a init_net callback after the list is updated.
	119	+ */
	120	+static DEFINE_XARRAY_FLAGS(rdma_nets, XA_FLAGS_ALLOC);
	121	+/*
	122	+ * rwsem to protect accessing the rdma_nets xarray entries.
	123	+ */
	124	+static DECLARE_RWSEM(rdma_nets_rwsem);
	125	+
	126	+bool ib_devices_shared_netns = true;
	127	+module_param_named(netns_mode, ib_devices_shared_netns, bool, 0444);
	128	+MODULE_PARM_DESC(netns_mode,
	129	+ "Share device among net namespaces; default=1 (shared)");
	130	+/**
	131	+ * rdma_dev_access_netns() - Return whether an rdma device can be accessed
	132	+ * from a specified net namespace or not.
	133	+ * @dev: Pointer to rdma device which needs to be checked
	134	+ * @net: Pointer to net namesapce for which access to be checked
	135	+ *
	136	+ * When the rdma device is in shared mode, it ignores the net namespace.
	137	+ * When the rdma device is exclusive to a net namespace, rdma device net
	138	+ * namespace is checked against the specified one.
	139	+ */
	140	+bool rdma_dev_access_netns(const struct ib_device dev, const struct net net)
	141	+{
	142	+ return (ib_devices_shared_netns \|\|
	143	+ net_eq(read_pnet(&dev->coredev.rdma_net), net));
	144	+}
	145	+EXPORT_SYMBOL(rdma_dev_access_netns);
	146	+
	147	+/*
	148	+ * xarray has this behavior where it won't iterate over NULL values stored in
	149	+ * allocated arrays. So we need our own iterator to see all values stored in
	150	+ * the array. This does the same thing as xa_for_each except that it also
	151	+ * returns NULL valued entries if the array is allocating. Simplified to only
	152	+ * work on simple xarrays.
	153	+ */
	154	+static void xan_find_marked(struct xarray xa, unsigned long *indexp,
	155	+ xa_mark_t filter)
	156	+{
	157	+ XA_STATE(xas, xa, *indexp);
	158	+ void *entry;
	159	+
	160	+ rcu_read_lock();
	161	+ do {
	162	+ entry = xas_find_marked(&xas, ULONG_MAX, filter);
	163	+ if (xa_is_zero(entry))
	164	+ break;
	165	+ } while (xas_retry(&xas, entry));
	166	+ rcu_read_unlock();
	167	+
	168	+ if (entry) {
	169	+ *indexp = xas.xa_index;
	170	+ if (xa_is_zero(entry))
	171	+ return NULL;
	172	+ return entry;
	173	+ }
	174	+ return XA_ERROR(-ENOENT);
	175	+}
	176	+#define xan_for_each_marked(xa, index, entry, filter) \
	177	+ for (index = 0, entry = xan_find_marked(xa, &(index), filter); \
	178	+ !xa_is_err(entry); \
	179	+ (index)++, entry = xan_find_marked(xa, &(index), filter))
	180	+
	181	+/* RCU hash table mapping netdevice pointers to struct ib_port_data */
	182	+static DEFINE_SPINLOCK(ndev_hash_lock);
	183	+static DECLARE_HASHTABLE(ndev_hash, 5);
	184	+
	185	+static void free_netdevs(struct ib_device *ib_dev);
	186	+static void ib_unregister_work(struct work_struct *work);
	187	+static void __ib_unregister_device(struct ib_device *device);
88	188	static int ib_security_change(struct notifier_block *nb, unsigned long event,
89	189	void *lsm_data);
90	190	static void ib_policy_change_task(struct work_struct *work);
91	191	static DECLARE_WORK(ib_policy_change_work, ib_policy_change_task);
92	192
	193	+static void __ibdev_printk(const char level, const struct ib_device ibdev,
	194	+ struct va_format *vaf)
	195	+{
	196	+ if (ibdev && ibdev->dev.parent)
	197	+ dev_printk_emit(level[1] - '0',
	198	+ ibdev->dev.parent,
	199	+ "%s %s %s: %pV",
	200	+ dev_driver_string(ibdev->dev.parent),
	201	+ dev_name(ibdev->dev.parent),
	202	+ dev_name(&ibdev->dev),
	203	+ vaf);
	204	+ else if (ibdev)
	205	+ printk("%s%s: %pV",
	206	+ level, dev_name(&ibdev->dev), vaf);
	207	+ else
	208	+ printk("%s(NULL ib_device): %pV", level, vaf);
	209	+}
	210	+
	211	+void ibdev_printk(const char level, const struct ib_device ibdev,
	212	+ const char *format, ...)
	213	+{
	214	+ struct va_format vaf;
	215	+ va_list args;
	216	+
	217	+ va_start(args, format);
	218	+
	219	+ vaf.fmt = format;
	220	+ vaf.va = &args;
	221	+
	222	+ __ibdev_printk(level, ibdev, &vaf);
	223	+
	224	+ va_end(args);
	225	+}
	226	+EXPORT_SYMBOL(ibdev_printk);
	227	+
	228	+#define define_ibdev_printk_level(func, level) \
	229	+void func(const struct ib_device ibdev, const char fmt, ...) \
	230	+{ \
	231	+ struct va_format vaf; \
	232	+ va_list args; \
	233	+ \
	234	+ va_start(args, fmt); \
	235	+ \
	236	+ vaf.fmt = fmt; \
	237	+ vaf.va = &args; \
	238	+ \
	239	+ __ibdev_printk(level, ibdev, &vaf); \
	240	+ \
	241	+ va_end(args); \
	242	+} \
	243	+EXPORT_SYMBOL(func);
	244	+
	245	+define_ibdev_printk_level(ibdev_emerg, KERN_EMERG);
	246	+define_ibdev_printk_level(ibdev_alert, KERN_ALERT);
	247	+define_ibdev_printk_level(ibdev_crit, KERN_CRIT);
	248	+define_ibdev_printk_level(ibdev_err, KERN_ERR);
	249	+define_ibdev_printk_level(ibdev_warn, KERN_WARNING);
	250	+define_ibdev_printk_level(ibdev_notice, KERN_NOTICE);
	251	+define_ibdev_printk_level(ibdev_info, KERN_INFO);
	252	+
93	253	static struct notifier_block ibdev_lsm_nb = {
94	254	.notifier_call = ib_security_change,
95	255	};
96	256
97		-static int ib_device_check_mandatory(struct ib_device *device)
	257	+static int rdma_dev_change_netns(struct ib_device device, struct net cur_net,
	258	+ struct net *net);
	259	+
	260	+/* Pointer to the RCU head at the start of the ib_port_data array */
	261	+struct ib_port_data_rcu {
	262	+ struct rcu_head rcu_head;
	263	+ struct ib_port_data pdata[];
	264	+};
	265	+
	266	+static void ib_device_check_mandatory(struct ib_device *device)
98	267	{
99		-#define IB_MANDATORY_FUNC(x) { offsetof(struct ib_device, x), #x }
	268	+#define IB_MANDATORY_FUNC(x) { offsetof(struct ib_device_ops, x), #x }
100	269	static const struct {
101	270	size_t offset;
102	271	char *name;
103	272	} mandatory_table[] = {
104	273	IB_MANDATORY_FUNC(query_device),
105	274	IB_MANDATORY_FUNC(query_port),
106		- IB_MANDATORY_FUNC(query_pkey),
107	275	IB_MANDATORY_FUNC(alloc_pd),
108	276	IB_MANDATORY_FUNC(dealloc_pd),
109	277	IB_MANDATORY_FUNC(create_qp),
..	..	@@ -121,110 +289,228 @@
121	289	};
122	290	int i;
123	291
	292	+ device->kverbs_provider = true;
124	293	for (i = 0; i < ARRAY_SIZE(mandatory_table); ++i) {
125		- if (!(void ) ((void ) device + mandatory_table[i].offset)) {
126		- pr_warn("Device %s is missing mandatory function %s\n",
127		- device->name, mandatory_table[i].name);
128		- return -EINVAL;
	294	+ if (!(void ) ((void ) &device->ops +
	295	+ mandatory_table[i].offset)) {
	296	+ device->kverbs_provider = false;
	297	+ break;
129	298	}
130	299	}
131		-
132		- return 0;
133		-}
134		-
135		-static struct ib_device *__ib_device_get_by_index(u32 index)
136		-{
137		- struct ib_device *device;
138		-
139		- list_for_each_entry(device, &device_list, core_list)
140		- if (device->index == index)
141		- return device;
142		-
143		- return NULL;
144	300	}
145	301
146	302	/*
147		- * Caller is responsible to return refrerence count by calling put_device()
	303	+ * Caller must perform ib_device_put() to return the device reference count
	304	+ * when ib_device_get_by_index() returns valid device pointer.
148	305	*/
149		-struct ib_device *ib_device_get_by_index(u32 index)
	306	+struct ib_device ib_device_get_by_index(const struct net net, u32 index)
150	307	{
151	308	struct ib_device *device;
152	309
153		- down_read(&lists_rwsem);
154		- device = __ib_device_get_by_index(index);
155		- if (device)
156		- get_device(&device->dev);
	310	+ down_read(&devices_rwsem);
	311	+ device = xa_load(&devices, index);
	312	+ if (device) {
	313	+ if (!rdma_dev_access_netns(device, net)) {
	314	+ device = NULL;
	315	+ goto out;
	316	+ }
157	317
158		- up_read(&lists_rwsem);
	318	+ if (!ib_device_try_get(device))
	319	+ device = NULL;
	320	+ }
	321	+out:
	322	+ up_read(&devices_rwsem);
159	323	return device;
160	324	}
	325	+
	326	+/**
	327	+ * ib_device_put - Release IB device reference
	328	+ * @device: device whose reference to be released
	329	+ *
	330	+ * ib_device_put() releases reference to the IB device to allow it to be
	331	+ * unregistered and eventually free.
	332	+ */
	333	+void ib_device_put(struct ib_device *device)
	334	+{
	335	+ if (refcount_dec_and_test(&device->refcount))
	336	+ complete(&device->unreg_completion);
	337	+}
	338	+EXPORT_SYMBOL(ib_device_put);
161	339
162	340	static struct ib_device __ib_device_get_by_name(const char name)
163	341	{
164	342	struct ib_device *device;
	343	+ unsigned long index;
165	344
166		- list_for_each_entry(device, &device_list, core_list)
167		- if (!strncmp(name, device->name, IB_DEVICE_NAME_MAX))
	345	+ xa_for_each (&devices, index, device)
	346	+ if (!strcmp(name, dev_name(&device->dev)))
168	347	return device;
169	348
170	349	return NULL;
171	350	}
172	351
173		-static int alloc_name(char *name)
	352	+/**
	353	+ * ib_device_get_by_name - Find an IB device by name
	354	+ * @name: The name to look for
	355	+ * @driver_id: The driver ID that must match (RDMA_DRIVER_UNKNOWN matches all)
	356	+ *
	357	+ * Find and hold an ib_device by its name. The caller must call
	358	+ * ib_device_put() on the returned pointer.
	359	+ */
	360	+struct ib_device ib_device_get_by_name(const char name,
	361	+ enum rdma_driver_id driver_id)
174	362	{
175		- unsigned long *inuse;
176		- char buf[IB_DEVICE_NAME_MAX];
177	363	struct ib_device *device;
178		- int i;
179	364
180		- inuse = (unsigned long *) get_zeroed_page(GFP_KERNEL);
181		- if (!inuse)
182		- return -ENOMEM;
	365	+ down_read(&devices_rwsem);
	366	+ device = __ib_device_get_by_name(name);
	367	+ if (device && driver_id != RDMA_DRIVER_UNKNOWN &&
	368	+ device->ops.driver_id != driver_id)
	369	+ device = NULL;
183	370
184		- list_for_each_entry(device, &device_list, core_list) {
185		- if (!sscanf(device->name, name, &i))
186		- continue;
187		- if (i < 0 \|\| i >= PAGE_SIZE * 8)
188		- continue;
189		- snprintf(buf, sizeof buf, name, i);
190		- if (!strncmp(buf, device->name, IB_DEVICE_NAME_MAX))
191		- set_bit(i, inuse);
	371	+ if (device) {
	372	+ if (!ib_device_try_get(device))
	373	+ device = NULL;
	374	+ }
	375	+ up_read(&devices_rwsem);
	376	+ return device;
	377	+}
	378	+EXPORT_SYMBOL(ib_device_get_by_name);
	379	+
	380	+static int rename_compat_devs(struct ib_device *device)
	381	+{
	382	+ struct ib_core_device *cdev;
	383	+ unsigned long index;
	384	+ int ret = 0;
	385	+
	386	+ mutex_lock(&device->compat_devs_mutex);
	387	+ xa_for_each (&device->compat_devs, index, cdev) {
	388	+ ret = device_rename(&cdev->dev, dev_name(&device->dev));
	389	+ if (ret) {
	390	+ dev_warn(&cdev->dev,
	391	+ "Fail to rename compatdev to new name %s\n",
	392	+ dev_name(&device->dev));
	393	+ break;
	394	+ }
	395	+ }
	396	+ mutex_unlock(&device->compat_devs_mutex);
	397	+ return ret;
	398	+}
	399	+
	400	+int ib_device_rename(struct ib_device ibdev, const char name)
	401	+{
	402	+ unsigned long index;
	403	+ void *client_data;
	404	+ int ret;
	405	+
	406	+ down_write(&devices_rwsem);
	407	+ if (!strcmp(name, dev_name(&ibdev->dev))) {
	408	+ up_write(&devices_rwsem);
	409	+ return 0;
192	410	}
193	411
194		- i = find_first_zero_bit(inuse, PAGE_SIZE * 8);
195		- free_page((unsigned long) inuse);
196		- snprintf(buf, sizeof buf, name, i);
	412	+ if (__ib_device_get_by_name(name)) {
	413	+ up_write(&devices_rwsem);
	414	+ return -EEXIST;
	415	+ }
197	416
198		- if (__ib_device_get_by_name(buf))
199		- return -ENFILE;
	417	+ ret = device_rename(&ibdev->dev, name);
	418	+ if (ret) {
	419	+ up_write(&devices_rwsem);
	420	+ return ret;
	421	+ }
200	422
201		- strlcpy(name, buf, IB_DEVICE_NAME_MAX);
	423	+ strlcpy(ibdev->name, name, IB_DEVICE_NAME_MAX);
	424	+ ret = rename_compat_devs(ibdev);
	425	+
	426	+ downgrade_write(&devices_rwsem);
	427	+ down_read(&ibdev->client_data_rwsem);
	428	+ xan_for_each_marked(&ibdev->client_data, index, client_data,
	429	+ CLIENT_DATA_REGISTERED) {
	430	+ struct ib_client *client = xa_load(&clients, index);
	431	+
	432	+ if (!client \|\| !client->rename)
	433	+ continue;
	434	+
	435	+ client->rename(ibdev, client_data);
	436	+ }
	437	+ up_read(&ibdev->client_data_rwsem);
	438	+ up_read(&devices_rwsem);
202	439	return 0;
	440	+}
	441	+
	442	+int ib_device_set_dim(struct ib_device *ibdev, u8 use_dim)
	443	+{
	444	+ if (use_dim > 1)
	445	+ return -EINVAL;
	446	+ ibdev->use_cq_dim = use_dim;
	447	+
	448	+ return 0;
	449	+}
	450	+
	451	+static int alloc_name(struct ib_device ibdev, const char name)
	452	+{
	453	+ struct ib_device *device;
	454	+ unsigned long index;
	455	+ struct ida inuse;
	456	+ int rc;
	457	+ int i;
	458	+
	459	+ lockdep_assert_held_write(&devices_rwsem);
	460	+ ida_init(&inuse);
	461	+ xa_for_each (&devices, index, device) {
	462	+ char buf[IB_DEVICE_NAME_MAX];
	463	+
	464	+ if (sscanf(dev_name(&device->dev), name, &i) != 1)
	465	+ continue;
	466	+ if (i < 0 \|\| i >= INT_MAX)
	467	+ continue;
	468	+ snprintf(buf, sizeof buf, name, i);
	469	+ if (strcmp(buf, dev_name(&device->dev)) != 0)
	470	+ continue;
	471	+
	472	+ rc = ida_alloc_range(&inuse, i, i, GFP_KERNEL);
	473	+ if (rc < 0)
	474	+ goto out;
	475	+ }
	476	+
	477	+ rc = ida_alloc(&inuse, GFP_KERNEL);
	478	+ if (rc < 0)
	479	+ goto out;
	480	+
	481	+ rc = dev_set_name(&ibdev->dev, name, rc);
	482	+out:
	483	+ ida_destroy(&inuse);
	484	+ return rc;
203	485	}
204	486
205	487	static void ib_device_release(struct device *device)
206	488	{
207	489	struct ib_device *dev = container_of(device, struct ib_device, dev);
208	490
209		- WARN_ON(dev->reg_state == IB_DEV_REGISTERED);
210		- if (dev->reg_state == IB_DEV_UNREGISTERED) {
211		- /*
212		- * In IB_DEV_UNINITIALIZED state, cache or port table
213		- * is not even created. Free cache and port table only when
214		- * device reaches UNREGISTERED state.
215		- */
	491	+ free_netdevs(dev);
	492	+ WARN_ON(refcount_read(&dev->refcount));
	493	+ if (dev->port_data) {
216	494	ib_cache_release_one(dev);
217		- kfree(dev->port_immutable);
	495	+ ib_security_release_port_pkey_list(dev);
	496	+ rdma_counter_release(dev);
	497	+ kfree_rcu(container_of(dev->port_data, struct ib_port_data_rcu,
	498	+ pdata[0]),
	499	+ rcu_head);
218	500	}
219		- kfree(dev);
	501	+
	502	+ mutex_destroy(&dev->unregistration_lock);
	503	+ mutex_destroy(&dev->compat_devs_mutex);
	504	+
	505	+ xa_destroy(&dev->compat_devs);
	506	+ xa_destroy(&dev->client_data);
	507	+ kfree_rcu(dev, rcu_head);
220	508	}
221	509
222	510	static int ib_device_uevent(struct device *device,
223	511	struct kobj_uevent_env *env)
224	512	{
225		- struct ib_device *dev = container_of(device, struct ib_device, dev);
226		-
227		- if (add_uevent_var(env, "NAME=%s", dev->name))
	513	+ if (add_uevent_var(env, "NAME=%s", dev_name(device)))
228	514	return -ENOMEM;
229	515
230	516	/*
..	..	@@ -234,14 +520,44 @@
234	520	return 0;
235	521	}
236	522
	523	+static const void net_namespace(struct device d)
	524	+{
	525	+ struct ib_core_device *coredev =
	526	+ container_of(d, struct ib_core_device, dev);
	527	+
	528	+ return read_pnet(&coredev->rdma_net);
	529	+}
	530	+
237	531	static struct class ib_class = {
238	532	.name = "infiniband",
239	533	.dev_release = ib_device_release,
240	534	.dev_uevent = ib_device_uevent,
	535	+ .ns_type = &net_ns_type_operations,
	536	+ .namespace = net_namespace,
241	537	};
242	538
	539	+static void rdma_init_coredev(struct ib_core_device *coredev,
	540	+ struct ib_device dev, struct net net)
	541	+{
	542	+ /* This BUILD_BUG_ON is intended to catch layout change
	543	+ * of union of ib_core_device and device.
	544	+ * dev must be the first element as ib_core and providers
	545	+ * driver uses it. Adding anything in ib_core_device before
	546	+ * device will break this assumption.
	547	+ */
	548	+ BUILD_BUG_ON(offsetof(struct ib_device, coredev.dev) !=
	549	+ offsetof(struct ib_device, dev));
	550	+
	551	+ coredev->dev.class = &ib_class;
	552	+ coredev->dev.groups = dev->groups;
	553	+ device_initialize(&coredev->dev);
	554	+ coredev->owner = dev;
	555	+ INIT_LIST_HEAD(&coredev->port_list);
	556	+ write_pnet(&coredev->rdma_net, net);
	557	+}
	558	+
243	559	/**
244		- * ib_alloc_device - allocate an IB device struct
	560	+ * _ib_alloc_device - allocate an IB device struct
245	561	* @size:size of structure to allocate
246	562	*
247	563	* Low-level drivers should use ib_alloc_device() to allocate &struct
..	..	@@ -250,7 +566,7 @@
250	566	* ib_dealloc_device() must be used to free structures allocated with
251	567	* ib_alloc_device().
252	568	*/
253		-struct ib_device *ib_alloc_device(size_t size)
	569	+struct ib_device *_ib_alloc_device(size_t size)
254	570	{
255	571	struct ib_device *device;
256	572
..	..	@@ -261,22 +577,32 @@
261	577	if (!device)
262	578	return NULL;
263	579
264		- rdma_restrack_init(&device->res);
	580	+ if (rdma_restrack_init(device)) {
	581	+ kfree(device);
	582	+ return NULL;
	583	+ }
265	584
266		- device->dev.class = &ib_class;
267		- device_initialize(&device->dev);
268		-
269		- dev_set_drvdata(&device->dev, device);
	585	+ device->groups[0] = &ib_dev_attr_group;
	586	+ rdma_init_coredev(&device->coredev, device, &init_net);
270	587
271	588	INIT_LIST_HEAD(&device->event_handler_list);
272		- spin_lock_init(&device->event_handler_lock);
273		- spin_lock_init(&device->client_data_lock);
274		- INIT_LIST_HEAD(&device->client_data_list);
275		- INIT_LIST_HEAD(&device->port_list);
	589	+ spin_lock_init(&device->qp_open_list_lock);
	590	+ init_rwsem(&device->event_handler_rwsem);
	591	+ mutex_init(&device->unregistration_lock);
	592	+ /*
	593	+ * client_data needs to be alloc because we don't want our mark to be
	594	+ * destroyed if the user stores NULL in the client data.
	595	+ */
	596	+ xa_init_flags(&device->client_data, XA_FLAGS_ALLOC);
	597	+ init_rwsem(&device->client_data_rwsem);
	598	+ xa_init_flags(&device->compat_devs, XA_FLAGS_ALLOC);
	599	+ mutex_init(&device->compat_devs_mutex);
	600	+ init_completion(&device->unreg_completion);
	601	+ INIT_WORK(&device->unregistration_work, ib_unregister_work);
276	602
277	603	return device;
278	604	}
279		-EXPORT_SYMBOL(ib_alloc_device);
	605	+EXPORT_SYMBOL(_ib_alloc_device);
280	606
281	607	/**
282	608	* ib_dealloc_device - free an IB device struct
..	..	@@ -286,32 +612,173 @@
286	612	*/
287	613	void ib_dealloc_device(struct ib_device *device)
288	614	{
289		- WARN_ON(device->reg_state != IB_DEV_UNREGISTERED &&
290		- device->reg_state != IB_DEV_UNINITIALIZED);
291		- rdma_restrack_clean(&device->res);
	615	+ if (device->ops.dealloc_driver)
	616	+ device->ops.dealloc_driver(device);
	617	+
	618	+ /*
	619	+ * ib_unregister_driver() requires all devices to remain in the xarray
	620	+ * while their ops are callable. The last op we call is dealloc_driver
	621	+ * above. This is needed to create a fence on op callbacks prior to
	622	+ * allowing the driver module to unload.
	623	+ */
	624	+ down_write(&devices_rwsem);
	625	+ if (xa_load(&devices, device->index) == device)
	626	+ xa_erase(&devices, device->index);
	627	+ up_write(&devices_rwsem);
	628	+
	629	+ /* Expedite releasing netdev references */
	630	+ free_netdevs(device);
	631	+
	632	+ WARN_ON(!xa_empty(&device->compat_devs));
	633	+ WARN_ON(!xa_empty(&device->client_data));
	634	+ WARN_ON(refcount_read(&device->refcount));
	635	+ rdma_restrack_clean(device);
	636	+ /* Balances with device_initialize */
292	637	put_device(&device->dev);
293	638	}
294	639	EXPORT_SYMBOL(ib_dealloc_device);
295	640
296		-static int add_client_context(struct ib_device device, struct ib_client client)
	641	+/*
	642	+ * add_client_context() and remove_client_context() must be safe against
	643	+ * parallel calls on the same device - registration/unregistration of both the
	644	+ * device and client can be occurring in parallel.
	645	+ *
	646	+ * The routines need to be a fence, any caller must not return until the add
	647	+ * or remove is fully completed.
	648	+ */
	649	+static int add_client_context(struct ib_device *device,
	650	+ struct ib_client *client)
297	651	{
298		- struct ib_client_data *context;
299		- unsigned long flags;
	652	+ int ret = 0;
300	653
301		- context = kmalloc(sizeof *context, GFP_KERNEL);
302		- if (!context)
	654	+ if (!device->kverbs_provider && !client->no_kverbs_req)
	655	+ return 0;
	656	+
	657	+ down_write(&device->client_data_rwsem);
	658	+ /*
	659	+ * So long as the client is registered hold both the client and device
	660	+ * unregistration locks.
	661	+ */
	662	+ if (!refcount_inc_not_zero(&client->uses))
	663	+ goto out_unlock;
	664	+ refcount_inc(&device->refcount);
	665	+
	666	+ /*
	667	+ * Another caller to add_client_context got here first and has already
	668	+ * completely initialized context.
	669	+ */
	670	+ if (xa_get_mark(&device->client_data, client->client_id,
	671	+ CLIENT_DATA_REGISTERED))
	672	+ goto out;
	673	+
	674	+ ret = xa_err(xa_store(&device->client_data, client->client_id, NULL,
	675	+ GFP_KERNEL));
	676	+ if (ret)
	677	+ goto out;
	678	+ downgrade_write(&device->client_data_rwsem);
	679	+ if (client->add) {
	680	+ if (client->add(device)) {
	681	+ /*
	682	+ * If a client fails to add then the error code is
	683	+ * ignored, but we won't call any more ops on this
	684	+ * client.
	685	+ */
	686	+ xa_erase(&device->client_data, client->client_id);
	687	+ up_read(&device->client_data_rwsem);
	688	+ ib_device_put(device);
	689	+ ib_client_put(client);
	690	+ return 0;
	691	+ }
	692	+ }
	693	+
	694	+ /* Readers shall not see a client until add has been completed */
	695	+ xa_set_mark(&device->client_data, client->client_id,
	696	+ CLIENT_DATA_REGISTERED);
	697	+ up_read(&device->client_data_rwsem);
	698	+ return 0;
	699	+
	700	+out:
	701	+ ib_device_put(device);
	702	+ ib_client_put(client);
	703	+out_unlock:
	704	+ up_write(&device->client_data_rwsem);
	705	+ return ret;
	706	+}
	707	+
	708	+static void remove_client_context(struct ib_device *device,
	709	+ unsigned int client_id)
	710	+{
	711	+ struct ib_client *client;
	712	+ void *client_data;
	713	+
	714	+ down_write(&device->client_data_rwsem);
	715	+ if (!xa_get_mark(&device->client_data, client_id,
	716	+ CLIENT_DATA_REGISTERED)) {
	717	+ up_write(&device->client_data_rwsem);
	718	+ return;
	719	+ }
	720	+ client_data = xa_load(&device->client_data, client_id);
	721	+ xa_clear_mark(&device->client_data, client_id, CLIENT_DATA_REGISTERED);
	722	+ client = xa_load(&clients, client_id);
	723	+ up_write(&device->client_data_rwsem);
	724	+
	725	+ /*
	726	+ * Notice we cannot be holding any exclusive locks when calling the
	727	+ * remove callback as the remove callback can recurse back into any
	728	+ * public functions in this module and thus try for any locks those
	729	+ * functions take.
	730	+ *
	731	+ * For this reason clients and drivers should not call the
	732	+ * unregistration functions will holdling any locks.
	733	+ */
	734	+ if (client->remove)
	735	+ client->remove(device, client_data);
	736	+
	737	+ xa_erase(&device->client_data, client_id);
	738	+ ib_device_put(device);
	739	+ ib_client_put(client);
	740	+}
	741	+
	742	+static int alloc_port_data(struct ib_device *device)
	743	+{
	744	+ struct ib_port_data_rcu *pdata_rcu;
	745	+ unsigned int port;
	746	+
	747	+ if (device->port_data)
	748	+ return 0;
	749	+
	750	+ /* This can only be called once the physical port range is defined */
	751	+ if (WARN_ON(!device->phys_port_cnt))
	752	+ return -EINVAL;
	753	+
	754	+ /*
	755	+ * device->port_data is indexed directly by the port number to make
	756	+ * access to this data as efficient as possible.
	757	+ *
	758	+ * Therefore port_data is declared as a 1 based array with potential
	759	+ * empty slots at the beginning.
	760	+ */
	761	+ pdata_rcu = kzalloc(struct_size(pdata_rcu, pdata,
	762	+ rdma_end_port(device) + 1),
	763	+ GFP_KERNEL);
	764	+ if (!pdata_rcu)
303	765	return -ENOMEM;
	766	+ /*
	767	+ * The rcu_head is put in front of the port data array and the stored
	768	+ * pointer is adjusted since we never need to see that member until
	769	+ * kfree_rcu.
	770	+ */
	771	+ device->port_data = pdata_rcu->pdata;
304	772
305		- context->client = client;
306		- context->data = NULL;
307		- context->going_down = false;
	773	+ rdma_for_each_port (device, port) {
	774	+ struct ib_port_data *pdata = &device->port_data[port];
308	775
309		- down_write(&lists_rwsem);
310		- spin_lock_irqsave(&device->client_data_lock, flags);
311		- list_add(&context->list, &device->client_data_list);
312		- spin_unlock_irqrestore(&device->client_data_lock, flags);
313		- up_write(&lists_rwsem);
314		-
	776	+ pdata->ib_dev = device;
	777	+ spin_lock_init(&pdata->pkey_list_lock);
	778	+ INIT_LIST_HEAD(&pdata->pkey_list);
	779	+ spin_lock_init(&pdata->netdev_lock);
	780	+ INIT_HLIST_NODE(&pdata->ndev_hash_link);
	781	+ }
315	782	return 0;
316	783	}
317	784
..	..	@@ -321,29 +788,20 @@
321	788	rdma_max_mad_size(dev, port) != 0);
322	789	}
323	790
324		-static int read_port_immutable(struct ib_device *device)
	791	+static int setup_port_data(struct ib_device *device)
325	792	{
	793	+ unsigned int port;
326	794	int ret;
327		- u8 start_port = rdma_start_port(device);
328		- u8 end_port = rdma_end_port(device);
329		- u8 port;
330	795
331		- /**
332		- * device->port_immutable is indexed directly by the port number to make
333		- * access to this data as efficient as possible.
334		- *
335		- * Therefore port_immutable is declared as a 1 based array with
336		- * potential empty slots at the beginning.
337		- */
338		- device->port_immutable = kcalloc(end_port + 1,
339		- sizeof(*device->port_immutable),
340		- GFP_KERNEL);
341		- if (!device->port_immutable)
342		- return -ENOMEM;
	796	+ ret = alloc_port_data(device);
	797	+ if (ret)
	798	+ return ret;
343	799
344		- for (port = start_port; port <= end_port; ++port) {
345		- ret = device->get_port_immutable(device, port,
346		- &device->port_immutable[port]);
	800	+ rdma_for_each_port (device, port) {
	801	+ struct ib_port_data *pdata = &device->port_data[port];
	802	+
	803	+ ret = device->ops.get_port_immutable(device, port,
	804	+ &pdata->immutable);
347	805	if (ret)
348	806	return ret;
349	807
..	..	@@ -355,46 +813,23 @@
355	813
356	814	void ib_get_device_fw_str(struct ib_device dev, char str)
357	815	{
358		- if (dev->get_dev_fw_str)
359		- dev->get_dev_fw_str(dev, str);
	816	+ if (dev->ops.get_dev_fw_str)
	817	+ dev->ops.get_dev_fw_str(dev, str);
360	818	else
361	819	str[0] = '\0';
362	820	}
363	821	EXPORT_SYMBOL(ib_get_device_fw_str);
364	822
365		-static int setup_port_pkey_list(struct ib_device *device)
366		-{
367		- int i;
368		-
369		- /**
370		- * device->port_pkey_list is indexed directly by the port number,
371		- * Therefore it is declared as a 1 based array with potential empty
372		- * slots at the beginning.
373		- */
374		- device->port_pkey_list = kcalloc(rdma_end_port(device) + 1,
375		- sizeof(*device->port_pkey_list),
376		- GFP_KERNEL);
377		-
378		- if (!device->port_pkey_list)
379		- return -ENOMEM;
380		-
381		- for (i = 0; i < (rdma_end_port(device) + 1); i++) {
382		- spin_lock_init(&device->port_pkey_list[i].list_lock);
383		- INIT_LIST_HEAD(&device->port_pkey_list[i].pkey_list);
384		- }
385		-
386		- return 0;
387		-}
388		-
389	823	static void ib_policy_change_task(struct work_struct *work)
390	824	{
391	825	struct ib_device *dev;
	826	+ unsigned long index;
392	827
393		- down_read(&lists_rwsem);
394		- list_for_each_entry(dev, &device_list, core_list) {
395		- int i;
	828	+ down_read(&devices_rwsem);
	829	+ xa_for_each_marked (&devices, index, dev, DEVICE_REGISTERED) {
	830	+ unsigned int i;
396	831
397		- for (i = rdma_start_port(dev); i <= rdma_end_port(dev); i++) {
	832	+ rdma_for_each_port (dev, i) {
398	833	u64 sp;
399	834	int ret = ib_get_cached_subnet_prefix(dev,
400	835	i,
..	..	@@ -407,7 +842,7 @@
407	842	ib_security_cache_change(dev, i, sp);
408	843	}
409	844	}
410		- up_read(&lists_rwsem);
	845	+ up_read(&devices_rwsem);
411	846	}
412	847
413	848	static int ib_security_change(struct notifier_block *nb, unsigned long event,
..	..	@@ -417,208 +852,857 @@
417	852	return NOTIFY_DONE;
418	853
419	854	schedule_work(&ib_policy_change_work);
	855	+ ib_mad_agent_security_change();
420	856
421	857	return NOTIFY_OK;
422	858	}
423	859
424		-/**
425		- * __dev_new_index - allocate an device index
426		- *
427		- * Returns a suitable unique value for a new device interface
428		- * number. It assumes that there are less than 2^32-1 ib devices
429		- * will be present in the system.
430		- */
431		-static u32 __dev_new_index(void)
	860	+static void compatdev_release(struct device *dev)
432	861	{
	862	+ struct ib_core_device *cdev =
	863	+ container_of(dev, struct ib_core_device, dev);
	864	+
	865	+ kfree(cdev);
	866	+}
	867	+
	868	+static int add_one_compat_dev(struct ib_device *device,
	869	+ struct rdma_dev_net *rnet)
	870	+{
	871	+ struct ib_core_device *cdev;
	872	+ int ret;
	873	+
	874	+ lockdep_assert_held(&rdma_nets_rwsem);
	875	+ if (!ib_devices_shared_netns)
	876	+ return 0;
	877	+
433	878	/*
434		- * The device index to allow stable naming.
435		- * Similar to struct net -> ifindex.
	879	+ * Create and add compat device in all namespaces other than where it
	880	+ * is currently bound to.
436	881	*/
437		- static u32 index;
	882	+ if (net_eq(read_pnet(&rnet->net),
	883	+ read_pnet(&device->coredev.rdma_net)))
	884	+ return 0;
438	885
439		- for (;;) {
440		- if (!(++index))
441		- index = 1;
442		-
443		- if (!__ib_device_get_by_index(index))
444		- return index;
	886	+ /*
	887	+ * The first of init_net() or ib_register_device() to take the
	888	+ * compat_devs_mutex wins and gets to add the device. Others will wait
	889	+ * for completion here.
	890	+ */
	891	+ mutex_lock(&device->compat_devs_mutex);
	892	+ cdev = xa_load(&device->compat_devs, rnet->id);
	893	+ if (cdev) {
	894	+ ret = 0;
	895	+ goto done;
445	896	}
	897	+ ret = xa_reserve(&device->compat_devs, rnet->id, GFP_KERNEL);
	898	+ if (ret)
	899	+ goto done;
	900	+
	901	+ cdev = kzalloc(sizeof(*cdev), GFP_KERNEL);
	902	+ if (!cdev) {
	903	+ ret = -ENOMEM;
	904	+ goto cdev_err;
	905	+ }
	906	+
	907	+ cdev->dev.parent = device->dev.parent;
	908	+ rdma_init_coredev(cdev, device, read_pnet(&rnet->net));
	909	+ cdev->dev.release = compatdev_release;
	910	+ ret = dev_set_name(&cdev->dev, "%s", dev_name(&device->dev));
	911	+ if (ret)
	912	+ goto add_err;
	913	+
	914	+ ret = device_add(&cdev->dev);
	915	+ if (ret)
	916	+ goto add_err;
	917	+ ret = ib_setup_port_attrs(cdev);
	918	+ if (ret)
	919	+ goto port_err;
	920	+
	921	+ ret = xa_err(xa_store(&device->compat_devs, rnet->id,
	922	+ cdev, GFP_KERNEL));
	923	+ if (ret)
	924	+ goto insert_err;
	925	+
	926	+ mutex_unlock(&device->compat_devs_mutex);
	927	+ return 0;
	928	+
	929	+insert_err:
	930	+ ib_free_port_attrs(cdev);
	931	+port_err:
	932	+ device_del(&cdev->dev);
	933	+add_err:
	934	+ put_device(&cdev->dev);
	935	+cdev_err:
	936	+ xa_release(&device->compat_devs, rnet->id);
	937	+done:
	938	+ mutex_unlock(&device->compat_devs_mutex);
	939	+ return ret;
	940	+}
	941	+
	942	+static void remove_one_compat_dev(struct ib_device *device, u32 id)
	943	+{
	944	+ struct ib_core_device *cdev;
	945	+
	946	+ mutex_lock(&device->compat_devs_mutex);
	947	+ cdev = xa_erase(&device->compat_devs, id);
	948	+ mutex_unlock(&device->compat_devs_mutex);
	949	+ if (cdev) {
	950	+ ib_free_port_attrs(cdev);
	951	+ device_del(&cdev->dev);
	952	+ put_device(&cdev->dev);
	953	+ }
	954	+}
	955	+
	956	+static void remove_compat_devs(struct ib_device *device)
	957	+{
	958	+ struct ib_core_device *cdev;
	959	+ unsigned long index;
	960	+
	961	+ xa_for_each (&device->compat_devs, index, cdev)
	962	+ remove_one_compat_dev(device, index);
	963	+}
	964	+
	965	+static int add_compat_devs(struct ib_device *device)
	966	+{
	967	+ struct rdma_dev_net *rnet;
	968	+ unsigned long index;
	969	+ int ret = 0;
	970	+
	971	+ lockdep_assert_held(&devices_rwsem);
	972	+
	973	+ down_read(&rdma_nets_rwsem);
	974	+ xa_for_each (&rdma_nets, index, rnet) {
	975	+ ret = add_one_compat_dev(device, rnet);
	976	+ if (ret)
	977	+ break;
	978	+ }
	979	+ up_read(&rdma_nets_rwsem);
	980	+ return ret;
	981	+}
	982	+
	983	+static void remove_all_compat_devs(void)
	984	+{
	985	+ struct ib_compat_device *cdev;
	986	+ struct ib_device *dev;
	987	+ unsigned long index;
	988	+
	989	+ down_read(&devices_rwsem);
	990	+ xa_for_each (&devices, index, dev) {
	991	+ unsigned long c_index = 0;
	992	+
	993	+ /* Hold nets_rwsem so that any other thread modifying this
	994	+ * system param can sync with this thread.
	995	+ */
	996	+ down_read(&rdma_nets_rwsem);
	997	+ xa_for_each (&dev->compat_devs, c_index, cdev)
	998	+ remove_one_compat_dev(dev, c_index);
	999	+ up_read(&rdma_nets_rwsem);
	1000	+ }
	1001	+ up_read(&devices_rwsem);
	1002	+}
	1003	+
	1004	+static int add_all_compat_devs(void)
	1005	+{
	1006	+ struct rdma_dev_net *rnet;
	1007	+ struct ib_device *dev;
	1008	+ unsigned long index;
	1009	+ int ret = 0;
	1010	+
	1011	+ down_read(&devices_rwsem);
	1012	+ xa_for_each_marked (&devices, index, dev, DEVICE_REGISTERED) {
	1013	+ unsigned long net_index = 0;
	1014	+
	1015	+ /* Hold nets_rwsem so that any other thread modifying this
	1016	+ * system param can sync with this thread.
	1017	+ */
	1018	+ down_read(&rdma_nets_rwsem);
	1019	+ xa_for_each (&rdma_nets, net_index, rnet) {
	1020	+ ret = add_one_compat_dev(dev, rnet);
	1021	+ if (ret)
	1022	+ break;
	1023	+ }
	1024	+ up_read(&rdma_nets_rwsem);
	1025	+ }
	1026	+ up_read(&devices_rwsem);
	1027	+ if (ret)
	1028	+ remove_all_compat_devs();
	1029	+ return ret;
	1030	+}
	1031	+
	1032	+int rdma_compatdev_set(u8 enable)
	1033	+{
	1034	+ struct rdma_dev_net *rnet;
	1035	+ unsigned long index;
	1036	+ int ret = 0;
	1037	+
	1038	+ down_write(&rdma_nets_rwsem);
	1039	+ if (ib_devices_shared_netns == enable) {
	1040	+ up_write(&rdma_nets_rwsem);
	1041	+ return 0;
	1042	+ }
	1043	+
	1044	+ /* enable/disable of compat devices is not supported
	1045	+ * when more than default init_net exists.
	1046	+ */
	1047	+ xa_for_each (&rdma_nets, index, rnet) {
	1048	+ ret++;
	1049	+ break;
	1050	+ }
	1051	+ if (!ret)
	1052	+ ib_devices_shared_netns = enable;
	1053	+ up_write(&rdma_nets_rwsem);
	1054	+ if (ret)
	1055	+ return -EBUSY;
	1056	+
	1057	+ if (enable)
	1058	+ ret = add_all_compat_devs();
	1059	+ else
	1060	+ remove_all_compat_devs();
	1061	+ return ret;
	1062	+}
	1063	+
	1064	+static void rdma_dev_exit_net(struct net *net)
	1065	+{
	1066	+ struct rdma_dev_net *rnet = rdma_net_to_dev_net(net);
	1067	+ struct ib_device *dev;
	1068	+ unsigned long index;
	1069	+ int ret;
	1070	+
	1071	+ down_write(&rdma_nets_rwsem);
	1072	+ /*
	1073	+ * Prevent the ID from being re-used and hide the id from xa_for_each.
	1074	+ */
	1075	+ ret = xa_err(xa_store(&rdma_nets, rnet->id, NULL, GFP_KERNEL));
	1076	+ WARN_ON(ret);
	1077	+ up_write(&rdma_nets_rwsem);
	1078	+
	1079	+ down_read(&devices_rwsem);
	1080	+ xa_for_each (&devices, index, dev) {
	1081	+ get_device(&dev->dev);
	1082	+ /*
	1083	+ * Release the devices_rwsem so that pontentially blocking
	1084	+ * device_del, doesn't hold the devices_rwsem for too long.
	1085	+ */
	1086	+ up_read(&devices_rwsem);
	1087	+
	1088	+ remove_one_compat_dev(dev, rnet->id);
	1089	+
	1090	+ /*
	1091	+ * If the real device is in the NS then move it back to init.
	1092	+ */
	1093	+ rdma_dev_change_netns(dev, net, &init_net);
	1094	+
	1095	+ put_device(&dev->dev);
	1096	+ down_read(&devices_rwsem);
	1097	+ }
	1098	+ up_read(&devices_rwsem);
	1099	+
	1100	+ rdma_nl_net_exit(rnet);
	1101	+ xa_erase(&rdma_nets, rnet->id);
	1102	+}
	1103	+
	1104	+static __net_init int rdma_dev_init_net(struct net *net)
	1105	+{
	1106	+ struct rdma_dev_net *rnet = rdma_net_to_dev_net(net);
	1107	+ unsigned long index;
	1108	+ struct ib_device *dev;
	1109	+ int ret;
	1110	+
	1111	+ write_pnet(&rnet->net, net);
	1112	+
	1113	+ ret = rdma_nl_net_init(rnet);
	1114	+ if (ret)
	1115	+ return ret;
	1116	+
	1117	+ /* No need to create any compat devices in default init_net. */
	1118	+ if (net_eq(net, &init_net))
	1119	+ return 0;
	1120	+
	1121	+ ret = xa_alloc(&rdma_nets, &rnet->id, rnet, xa_limit_32b, GFP_KERNEL);
	1122	+ if (ret) {
	1123	+ rdma_nl_net_exit(rnet);
	1124	+ return ret;
	1125	+ }
	1126	+
	1127	+ down_read(&devices_rwsem);
	1128	+ xa_for_each_marked (&devices, index, dev, DEVICE_REGISTERED) {
	1129	+ /* Hold nets_rwsem so that netlink command cannot change
	1130	+ * system configuration for device sharing mode.
	1131	+ */
	1132	+ down_read(&rdma_nets_rwsem);
	1133	+ ret = add_one_compat_dev(dev, rnet);
	1134	+ up_read(&rdma_nets_rwsem);
	1135	+ if (ret)
	1136	+ break;
	1137	+ }
	1138	+ up_read(&devices_rwsem);
	1139	+
	1140	+ if (ret)
	1141	+ rdma_dev_exit_net(net);
	1142	+
	1143	+ return ret;
	1144	+}
	1145	+
	1146	+/*
	1147	+ * Assign the unique string device name and the unique device index. This is
	1148	+ * undone by ib_dealloc_device.
	1149	+ */
	1150	+static int assign_name(struct ib_device device, const char name)
	1151	+{
	1152	+ static u32 last_id;
	1153	+ int ret;
	1154	+
	1155	+ down_write(&devices_rwsem);
	1156	+ /* Assign a unique name to the device */
	1157	+ if (strchr(name, '%'))
	1158	+ ret = alloc_name(device, name);
	1159	+ else
	1160	+ ret = dev_set_name(&device->dev, name);
	1161	+ if (ret)
	1162	+ goto out;
	1163	+
	1164	+ if (__ib_device_get_by_name(dev_name(&device->dev))) {
	1165	+ ret = -ENFILE;
	1166	+ goto out;
	1167	+ }
	1168	+ strlcpy(device->name, dev_name(&device->dev), IB_DEVICE_NAME_MAX);
	1169	+
	1170	+ ret = xa_alloc_cyclic(&devices, &device->index, device, xa_limit_31b,
	1171	+ &last_id, GFP_KERNEL);
	1172	+ if (ret > 0)
	1173	+ ret = 0;
	1174	+
	1175	+out:
	1176	+ up_write(&devices_rwsem);
	1177	+ return ret;
	1178	+}
	1179	+
	1180	+/*
	1181	+ * setup_device() allocates memory and sets up data that requires calling the
	1182	+ * device ops, this is the only reason these actions are not done during
	1183	+ * ib_alloc_device. It is undone by ib_dealloc_device().
	1184	+ */
	1185	+static int setup_device(struct ib_device *device)
	1186	+{
	1187	+ struct ib_udata uhw = {.outlen = 0, .inlen = 0};
	1188	+ int ret;
	1189	+
	1190	+ ib_device_check_mandatory(device);
	1191	+
	1192	+ ret = setup_port_data(device);
	1193	+ if (ret) {
	1194	+ dev_warn(&device->dev, "Couldn't create per-port data\n");
	1195	+ return ret;
	1196	+ }
	1197	+
	1198	+ memset(&device->attrs, 0, sizeof(device->attrs));
	1199	+ ret = device->ops.query_device(device, &device->attrs, &uhw);
	1200	+ if (ret) {
	1201	+ dev_warn(&device->dev,
	1202	+ "Couldn't query the device attributes\n");
	1203	+ return ret;
	1204	+ }
	1205	+
	1206	+ return 0;
	1207	+}
	1208	+
	1209	+static void disable_device(struct ib_device *device)
	1210	+{
	1211	+ u32 cid;
	1212	+
	1213	+ WARN_ON(!refcount_read(&device->refcount));
	1214	+
	1215	+ down_write(&devices_rwsem);
	1216	+ xa_clear_mark(&devices, device->index, DEVICE_REGISTERED);
	1217	+ up_write(&devices_rwsem);
	1218	+
	1219	+ /*
	1220	+ * Remove clients in LIFO order, see assign_client_id. This could be
	1221	+ * more efficient if xarray learns to reverse iterate. Since no new
	1222	+ * clients can be added to this ib_device past this point we only need
	1223	+ * the maximum possible client_id value here.
	1224	+ */
	1225	+ down_read(&clients_rwsem);
	1226	+ cid = highest_client_id;
	1227	+ up_read(&clients_rwsem);
	1228	+ while (cid) {
	1229	+ cid--;
	1230	+ remove_client_context(device, cid);
	1231	+ }
	1232	+
	1233	+ ib_cq_pool_destroy(device);
	1234	+
	1235	+ /* Pairs with refcount_set in enable_device */
	1236	+ ib_device_put(device);
	1237	+ wait_for_completion(&device->unreg_completion);
	1238	+
	1239	+ /*
	1240	+ * compat devices must be removed after device refcount drops to zero.
	1241	+ * Otherwise init_net() may add more compatdevs after removing compat
	1242	+ * devices and before device is disabled.
	1243	+ */
	1244	+ remove_compat_devs(device);
	1245	+}
	1246	+
	1247	+/*
	1248	+ * An enabled device is visible to all clients and to all the public facing
	1249	+ * APIs that return a device pointer. This always returns with a new get, even
	1250	+ * if it fails.
	1251	+ */
	1252	+static int enable_device_and_get(struct ib_device *device)
	1253	+{
	1254	+ struct ib_client *client;
	1255	+ unsigned long index;
	1256	+ int ret = 0;
	1257	+
	1258	+ /*
	1259	+ * One ref belongs to the xa and the other belongs to this
	1260	+ * thread. This is needed to guard against parallel unregistration.
	1261	+ */
	1262	+ refcount_set(&device->refcount, 2);
	1263	+ down_write(&devices_rwsem);
	1264	+ xa_set_mark(&devices, device->index, DEVICE_REGISTERED);
	1265	+
	1266	+ /*
	1267	+ * By using downgrade_write() we ensure that no other thread can clear
	1268	+ * DEVICE_REGISTERED while we are completing the client setup.
	1269	+ */
	1270	+ downgrade_write(&devices_rwsem);
	1271	+
	1272	+ if (device->ops.enable_driver) {
	1273	+ ret = device->ops.enable_driver(device);
	1274	+ if (ret)
	1275	+ goto out;
	1276	+ }
	1277	+
	1278	+ ib_cq_pool_init(device);
	1279	+
	1280	+ down_read(&clients_rwsem);
	1281	+ xa_for_each_marked (&clients, index, client, CLIENT_REGISTERED) {
	1282	+ ret = add_client_context(device, client);
	1283	+ if (ret)
	1284	+ break;
	1285	+ }
	1286	+ up_read(&clients_rwsem);
	1287	+ if (!ret)
	1288	+ ret = add_compat_devs(device);
	1289	+out:
	1290	+ up_read(&devices_rwsem);
	1291	+ return ret;
	1292	+}
	1293	+
	1294	+static void prevent_dealloc_device(struct ib_device *ib_dev)
	1295	+{
446	1296	}
447	1297
448	1298	/**
449	1299	* ib_register_device - Register an IB device with IB core
450		- * @device:Device to register
	1300	+ * @device: Device to register
	1301	+ * @name: unique string device name. This may include a '%' which will
	1302	+ * cause a unique index to be added to the passed device name.
	1303	+ * @dma_device: pointer to a DMA-capable device. If %NULL, then the IB
	1304	+ * device will be used. In this case the caller should fully
	1305	+ * setup the ibdev for DMA. This usually means using dma_virt_ops.
451	1306	*
452	1307	* Low-level drivers use ib_register_device() to register their
453	1308	* devices with the IB core. All registered clients will receive a
454	1309	* callback for each device that is added. @device must be allocated
455	1310	* with ib_alloc_device().
	1311	+ *
	1312	+ * If the driver uses ops.dealloc_driver and calls any ib_unregister_device()
	1313	+ * asynchronously then the device pointer may become freed as soon as this
	1314	+ * function returns.
456	1315	*/
457		-int ib_register_device(struct ib_device *device,
458		- int (port_callback)(struct ib_device ,
459		- u8, struct kobject *))
	1316	+int ib_register_device(struct ib_device device, const char name,
	1317	+ struct device *dma_device)
460	1318	{
461	1319	int ret;
462		- struct ib_client *client;
463		- struct ib_udata uhw = {.outlen = 0, .inlen = 0};
464		- struct device *parent = device->dev.parent;
465	1320
466		- WARN_ON_ONCE(device->dma_device);
467		- if (device->dev.dma_ops) {
468		- /*
469		- * The caller provided custom DMA operations. Copy the
470		- * DMA-related fields that are used by e.g. dma_alloc_coherent()
471		- * into device->dev.
472		- */
473		- device->dma_device = &device->dev;
474		- if (!device->dev.dma_mask) {
475		- if (parent)
476		- device->dev.dma_mask = parent->dma_mask;
477		- else
478		- WARN_ON_ONCE(true);
479		- }
480		- if (!device->dev.coherent_dma_mask) {
481		- if (parent)
482		- device->dev.coherent_dma_mask =
483		- parent->coherent_dma_mask;
484		- else
485		- WARN_ON_ONCE(true);
486		- }
487		- } else {
488		- /*
489		- * The caller did not provide custom DMA operations. Use the
490		- * DMA mapping operations of the parent device.
491		- */
492		- WARN_ON_ONCE(!parent);
493		- device->dma_device = parent;
494		- }
	1321	+ ret = assign_name(device, name);
	1322	+ if (ret)
	1323	+ return ret;
495	1324
496		- mutex_lock(&device_mutex);
	1325	+ /*
	1326	+ * If the caller does not provide a DMA capable device then the IB core
	1327	+ * will set up ib_sge and scatterlist structures that stash the kernel
	1328	+ * virtual address into the address field.
	1329	+ */
	1330	+ WARN_ON(dma_device && !dma_device->dma_parms);
	1331	+ device->dma_device = dma_device;
497	1332
498		- if (strchr(device->name, '%')) {
499		- ret = alloc_name(device->name);
500		- if (ret)
501		- goto out;
502		- }
503		-
504		- if (ib_device_check_mandatory(device)) {
505		- ret = -EINVAL;
506		- goto out;
507		- }
508		-
509		- ret = read_port_immutable(device);
510		- if (ret) {
511		- pr_warn("Couldn't create per port immutable data %s\n",
512		- device->name);
513		- goto out;
514		- }
515		-
516		- ret = setup_port_pkey_list(device);
517		- if (ret) {
518		- pr_warn("Couldn't create per port_pkey_list\n");
519		- goto out;
520		- }
	1333	+ ret = setup_device(device);
	1334	+ if (ret)
	1335	+ return ret;
521	1336
522	1337	ret = ib_cache_setup_one(device);
523	1338	if (ret) {
524		- pr_warn("Couldn't set up InfiniBand P_Key/GID cache\n");
525		- goto port_cleanup;
	1339	+ dev_warn(&device->dev,
	1340	+ "Couldn't set up InfiniBand P_Key/GID cache\n");
	1341	+ return ret;
526	1342	}
527	1343
528		- ret = ib_device_register_rdmacg(device);
529		- if (ret) {
530		- pr_warn("Couldn't register device with rdma cgroup\n");
531		- goto cache_cleanup;
532		- }
	1344	+ ib_device_register_rdmacg(device);
533	1345
534		- memset(&device->attrs, 0, sizeof(device->attrs));
535		- ret = device->query_device(device, &device->attrs, &uhw);
536		- if (ret) {
537		- pr_warn("Couldn't query the device attributes\n");
	1346	+ rdma_counter_init(device);
	1347	+
	1348	+ /*
	1349	+ * Ensure that ADD uevent is not fired because it
	1350	+ * is too early amd device is not initialized yet.
	1351	+ */
	1352	+ dev_set_uevent_suppress(&device->dev, true);
	1353	+ ret = device_add(&device->dev);
	1354	+ if (ret)
538	1355	goto cg_cleanup;
539		- }
540	1356
541		- ret = ib_device_register_sysfs(device, port_callback);
	1357	+ ret = ib_device_register_sysfs(device);
542	1358	if (ret) {
543		- pr_warn("Couldn't register device %s with driver model\n",
544		- device->name);
545		- goto cg_cleanup;
	1359	+ dev_warn(&device->dev,
	1360	+ "Couldn't register device with driver model\n");
	1361	+ goto dev_cleanup;
546	1362	}
547	1363
548		- device->reg_state = IB_DEV_REGISTERED;
	1364	+ ret = enable_device_and_get(device);
	1365	+ if (ret) {
	1366	+ void (dealloc_fn)(struct ib_device );
549	1367
550		- list_for_each_entry(client, &client_list, list)
551		- if (!add_client_context(device, client) && client->add)
552		- client->add(device);
	1368	+ /*
	1369	+ * If we hit this error flow then we don't want to
	1370	+ * automatically dealloc the device since the caller is
	1371	+ * expected to call ib_dealloc_device() after
	1372	+ * ib_register_device() fails. This is tricky due to the
	1373	+ * possibility for a parallel unregistration along with this
	1374	+ * error flow. Since we have a refcount here we know any
	1375	+ * parallel flow is stopped in disable_device and will see the
	1376	+ * special dealloc_driver pointer, causing the responsibility to
	1377	+ * ib_dealloc_device() to revert back to this thread.
	1378	+ */
	1379	+ dealloc_fn = device->ops.dealloc_driver;
	1380	+ device->ops.dealloc_driver = prevent_dealloc_device;
	1381	+ ib_device_put(device);
	1382	+ __ib_unregister_device(device);
	1383	+ device->ops.dealloc_driver = dealloc_fn;
	1384	+ dev_set_uevent_suppress(&device->dev, false);
	1385	+ return ret;
	1386	+ }
	1387	+ dev_set_uevent_suppress(&device->dev, false);
	1388	+ /* Mark for userspace that device is ready */
	1389	+ kobject_uevent(&device->dev.kobj, KOBJ_ADD);
	1390	+ ib_device_put(device);
553	1391
554		- device->index = __dev_new_index();
555		- down_write(&lists_rwsem);
556		- list_add_tail(&device->core_list, &device_list);
557		- up_write(&lists_rwsem);
558		- mutex_unlock(&device_mutex);
559	1392	return 0;
560	1393
	1394	+dev_cleanup:
	1395	+ device_del(&device->dev);
561	1396	cg_cleanup:
	1397	+ dev_set_uevent_suppress(&device->dev, false);
562	1398	ib_device_unregister_rdmacg(device);
563		-cache_cleanup:
564	1399	ib_cache_cleanup_one(device);
565		- ib_cache_release_one(device);
566		-port_cleanup:
567		- kfree(device->port_immutable);
568		-out:
569		- mutex_unlock(&device_mutex);
570	1400	return ret;
571	1401	}
572	1402	EXPORT_SYMBOL(ib_register_device);
573	1403
	1404	+/* Callers must hold a get on the device. */
	1405	+static void __ib_unregister_device(struct ib_device *ib_dev)
	1406	+{
	1407	+ /*
	1408	+ * We have a registration lock so that all the calls to unregister are
	1409	+ * fully fenced, once any unregister returns the device is truely
	1410	+ * unregistered even if multiple callers are unregistering it at the
	1411	+ * same time. This also interacts with the registration flow and
	1412	+ * provides sane semantics if register and unregister are racing.
	1413	+ */
	1414	+ mutex_lock(&ib_dev->unregistration_lock);
	1415	+ if (!refcount_read(&ib_dev->refcount))
	1416	+ goto out;
	1417	+
	1418	+ disable_device(ib_dev);
	1419	+
	1420	+ /* Expedite removing unregistered pointers from the hash table */
	1421	+ free_netdevs(ib_dev);
	1422	+
	1423	+ ib_device_unregister_sysfs(ib_dev);
	1424	+ device_del(&ib_dev->dev);
	1425	+ ib_device_unregister_rdmacg(ib_dev);
	1426	+ ib_cache_cleanup_one(ib_dev);
	1427	+
	1428	+ /*
	1429	+ * Drivers using the new flow may not call ib_dealloc_device except
	1430	+ * in error unwind prior to registration success.
	1431	+ */
	1432	+ if (ib_dev->ops.dealloc_driver &&
	1433	+ ib_dev->ops.dealloc_driver != prevent_dealloc_device) {
	1434	+ WARN_ON(kref_read(&ib_dev->dev.kobj.kref) <= 1);
	1435	+ ib_dealloc_device(ib_dev);
	1436	+ }
	1437	+out:
	1438	+ mutex_unlock(&ib_dev->unregistration_lock);
	1439	+}
	1440	+
574	1441	/**
575	1442	* ib_unregister_device - Unregister an IB device
576		- * @device:Device to unregister
	1443	+ * @ib_dev: The device to unregister
577	1444	*
578	1445	* Unregister an IB device. All clients will receive a remove callback.
	1446	+ *
	1447	+ * Callers should call this routine only once, and protect against races with
	1448	+ * registration. Typically it should only be called as part of a remove
	1449	+ * callback in an implementation of driver core's struct device_driver and
	1450	+ * related.
	1451	+ *
	1452	+ * If ops.dealloc_driver is used then ib_dev will be freed upon return from
	1453	+ * this function.
579	1454	*/
580		-void ib_unregister_device(struct ib_device *device)
	1455	+void ib_unregister_device(struct ib_device *ib_dev)
581	1456	{
582		- struct ib_client_data context, tmp;
583		- unsigned long flags;
584		-
585		- mutex_lock(&device_mutex);
586		-
587		- down_write(&lists_rwsem);
588		- list_del(&device->core_list);
589		- spin_lock_irqsave(&device->client_data_lock, flags);
590		- list_for_each_entry_safe(context, tmp, &device->client_data_list, list)
591		- context->going_down = true;
592		- spin_unlock_irqrestore(&device->client_data_lock, flags);
593		- downgrade_write(&lists_rwsem);
594		-
595		- list_for_each_entry_safe(context, tmp, &device->client_data_list,
596		- list) {
597		- if (context->client->remove)
598		- context->client->remove(device, context->data);
599		- }
600		- up_read(&lists_rwsem);
601		-
602		- ib_device_unregister_sysfs(device);
603		- ib_device_unregister_rdmacg(device);
604		-
605		- mutex_unlock(&device_mutex);
606		-
607		- ib_cache_cleanup_one(device);
608		-
609		- ib_security_destroy_port_pkey_list(device);
610		- kfree(device->port_pkey_list);
611		-
612		- down_write(&lists_rwsem);
613		- spin_lock_irqsave(&device->client_data_lock, flags);
614		- list_for_each_entry_safe(context, tmp, &device->client_data_list, list)
615		- kfree(context);
616		- spin_unlock_irqrestore(&device->client_data_lock, flags);
617		- up_write(&lists_rwsem);
618		-
619		- device->reg_state = IB_DEV_UNREGISTERED;
	1457	+ get_device(&ib_dev->dev);
	1458	+ __ib_unregister_device(ib_dev);
	1459	+ put_device(&ib_dev->dev);
620	1460	}
621	1461	EXPORT_SYMBOL(ib_unregister_device);
	1462	+
	1463	+/**
	1464	+ * ib_unregister_device_and_put - Unregister a device while holding a 'get'
	1465	+ * @ib_dev: The device to unregister
	1466	+ *
	1467	+ * This is the same as ib_unregister_device(), except it includes an internal
	1468	+ * ib_device_put() that should match a 'get' obtained by the caller.
	1469	+ *
	1470	+ * It is safe to call this routine concurrently from multiple threads while
	1471	+ * holding the 'get'. When the function returns the device is fully
	1472	+ * unregistered.
	1473	+ *
	1474	+ * Drivers using this flow MUST use the driver_unregister callback to clean up
	1475	+ * their resources associated with the device and dealloc it.
	1476	+ */
	1477	+void ib_unregister_device_and_put(struct ib_device *ib_dev)
	1478	+{
	1479	+ WARN_ON(!ib_dev->ops.dealloc_driver);
	1480	+ get_device(&ib_dev->dev);
	1481	+ ib_device_put(ib_dev);
	1482	+ __ib_unregister_device(ib_dev);
	1483	+ put_device(&ib_dev->dev);
	1484	+}
	1485	+EXPORT_SYMBOL(ib_unregister_device_and_put);
	1486	+
	1487	+/**
	1488	+ * ib_unregister_driver - Unregister all IB devices for a driver
	1489	+ * @driver_id: The driver to unregister
	1490	+ *
	1491	+ * This implements a fence for device unregistration. It only returns once all
	1492	+ * devices associated with the driver_id have fully completed their
	1493	+ * unregistration and returned from ib_unregister_device*().
	1494	+ *
	1495	+ * If device's are not yet unregistered it goes ahead and starts unregistering
	1496	+ * them.
	1497	+ *
	1498	+ * This does not block creation of new devices with the given driver_id, that
	1499	+ * is the responsibility of the caller.
	1500	+ */
	1501	+void ib_unregister_driver(enum rdma_driver_id driver_id)
	1502	+{
	1503	+ struct ib_device *ib_dev;
	1504	+ unsigned long index;
	1505	+
	1506	+ down_read(&devices_rwsem);
	1507	+ xa_for_each (&devices, index, ib_dev) {
	1508	+ if (ib_dev->ops.driver_id != driver_id)
	1509	+ continue;
	1510	+
	1511	+ get_device(&ib_dev->dev);
	1512	+ up_read(&devices_rwsem);
	1513	+
	1514	+ WARN_ON(!ib_dev->ops.dealloc_driver);
	1515	+ __ib_unregister_device(ib_dev);
	1516	+
	1517	+ put_device(&ib_dev->dev);
	1518	+ down_read(&devices_rwsem);
	1519	+ }
	1520	+ up_read(&devices_rwsem);
	1521	+}
	1522	+EXPORT_SYMBOL(ib_unregister_driver);
	1523	+
	1524	+static void ib_unregister_work(struct work_struct *work)
	1525	+{
	1526	+ struct ib_device *ib_dev =
	1527	+ container_of(work, struct ib_device, unregistration_work);
	1528	+
	1529	+ __ib_unregister_device(ib_dev);
	1530	+ put_device(&ib_dev->dev);
	1531	+}
	1532	+
	1533	+/**
	1534	+ * ib_unregister_device_queued - Unregister a device using a work queue
	1535	+ * @ib_dev: The device to unregister
	1536	+ *
	1537	+ * This schedules an asynchronous unregistration using a WQ for the device. A
	1538	+ * driver should use this to avoid holding locks while doing unregistration,
	1539	+ * such as holding the RTNL lock.
	1540	+ *
	1541	+ * Drivers using this API must use ib_unregister_driver before module unload
	1542	+ * to ensure that all scheduled unregistrations have completed.
	1543	+ */
	1544	+void ib_unregister_device_queued(struct ib_device *ib_dev)
	1545	+{
	1546	+ WARN_ON(!refcount_read(&ib_dev->refcount));
	1547	+ WARN_ON(!ib_dev->ops.dealloc_driver);
	1548	+ get_device(&ib_dev->dev);
	1549	+ if (!queue_work(system_unbound_wq, &ib_dev->unregistration_work))
	1550	+ put_device(&ib_dev->dev);
	1551	+}
	1552	+EXPORT_SYMBOL(ib_unregister_device_queued);
	1553	+
	1554	+/*
	1555	+ * The caller must pass in a device that has the kref held and the refcount
	1556	+ * released. If the device is in cur_net and still registered then it is moved
	1557	+ * into net.
	1558	+ */
	1559	+static int rdma_dev_change_netns(struct ib_device device, struct net cur_net,
	1560	+ struct net *net)
	1561	+{
	1562	+ int ret2 = -EINVAL;
	1563	+ int ret;
	1564	+
	1565	+ mutex_lock(&device->unregistration_lock);
	1566	+
	1567	+ /*
	1568	+ * If a device not under ib_device_get() or if the unregistration_lock
	1569	+ * is not held, the namespace can be changed, or it can be unregistered.
	1570	+ * Check again under the lock.
	1571	+ */
	1572	+ if (refcount_read(&device->refcount) == 0 \|\|
	1573	+ !net_eq(cur_net, read_pnet(&device->coredev.rdma_net))) {
	1574	+ ret = -ENODEV;
	1575	+ goto out;
	1576	+ }
	1577	+
	1578	+ kobject_uevent(&device->dev.kobj, KOBJ_REMOVE);
	1579	+ disable_device(device);
	1580	+
	1581	+ /*
	1582	+ * At this point no one can be using the device, so it is safe to
	1583	+ * change the namespace.
	1584	+ */
	1585	+ write_pnet(&device->coredev.rdma_net, net);
	1586	+
	1587	+ down_read(&devices_rwsem);
	1588	+ /*
	1589	+ * Currently rdma devices are system wide unique. So the device name
	1590	+ * is guaranteed free in the new namespace. Publish the new namespace
	1591	+ * at the sysfs level.
	1592	+ */
	1593	+ ret = device_rename(&device->dev, dev_name(&device->dev));
	1594	+ up_read(&devices_rwsem);
	1595	+ if (ret) {
	1596	+ dev_warn(&device->dev,
	1597	+ "%s: Couldn't rename device after namespace change\n",
	1598	+ __func__);
	1599	+ /* Try and put things back and re-enable the device */
	1600	+ write_pnet(&device->coredev.rdma_net, cur_net);
	1601	+ }
	1602	+
	1603	+ ret2 = enable_device_and_get(device);
	1604	+ if (ret2) {
	1605	+ /*
	1606	+ * This shouldn't really happen, but if it does, let the user
	1607	+ * retry at later point. So don't disable the device.
	1608	+ */
	1609	+ dev_warn(&device->dev,
	1610	+ "%s: Couldn't re-enable device after namespace change\n",
	1611	+ __func__);
	1612	+ }
	1613	+ kobject_uevent(&device->dev.kobj, KOBJ_ADD);
	1614	+
	1615	+ ib_device_put(device);
	1616	+out:
	1617	+ mutex_unlock(&device->unregistration_lock);
	1618	+ if (ret)
	1619	+ return ret;
	1620	+ return ret2;
	1621	+}
	1622	+
	1623	+int ib_device_set_netns_put(struct sk_buff *skb,
	1624	+ struct ib_device *dev, u32 ns_fd)
	1625	+{
	1626	+ struct net *net;
	1627	+ int ret;
	1628	+
	1629	+ net = get_net_ns_by_fd(ns_fd);
	1630	+ if (IS_ERR(net)) {
	1631	+ ret = PTR_ERR(net);
	1632	+ goto net_err;
	1633	+ }
	1634	+
	1635	+ if (!netlink_ns_capable(skb, net->user_ns, CAP_NET_ADMIN)) {
	1636	+ ret = -EPERM;
	1637	+ goto ns_err;
	1638	+ }
	1639	+
	1640	+ /*
	1641	+ * Currently supported only for those providers which support
	1642	+ * disassociation and don't do port specific sysfs init. Once a
	1643	+ * port_cleanup infrastructure is implemented, this limitation will be
	1644	+ * removed.
	1645	+ */
	1646	+ if (!dev->ops.disassociate_ucontext \|\| dev->ops.init_port \|\|
	1647	+ ib_devices_shared_netns) {
	1648	+ ret = -EOPNOTSUPP;
	1649	+ goto ns_err;
	1650	+ }
	1651	+
	1652	+ get_device(&dev->dev);
	1653	+ ib_device_put(dev);
	1654	+ ret = rdma_dev_change_netns(dev, current->nsproxy->net_ns, net);
	1655	+ put_device(&dev->dev);
	1656	+
	1657	+ put_net(net);
	1658	+ return ret;
	1659	+
	1660	+ns_err:
	1661	+ put_net(net);
	1662	+net_err:
	1663	+ ib_device_put(dev);
	1664	+ return ret;
	1665	+}
	1666	+
	1667	+static struct pernet_operations rdma_dev_net_ops = {
	1668	+ .init = rdma_dev_init_net,
	1669	+ .exit = rdma_dev_exit_net,
	1670	+ .id = &rdma_dev_net_id,
	1671	+ .size = sizeof(struct rdma_dev_net),
	1672	+};
	1673	+
	1674	+static int assign_client_id(struct ib_client *client)
	1675	+{
	1676	+ int ret;
	1677	+
	1678	+ down_write(&clients_rwsem);
	1679	+ /*
	1680	+ * The add/remove callbacks must be called in FIFO/LIFO order. To
	1681	+ * achieve this we assign client_ids so they are sorted in
	1682	+ * registration order.
	1683	+ */
	1684	+ client->client_id = highest_client_id;
	1685	+ ret = xa_insert(&clients, client->client_id, client, GFP_KERNEL);
	1686	+ if (ret)
	1687	+ goto out;
	1688	+
	1689	+ highest_client_id++;
	1690	+ xa_set_mark(&clients, client->client_id, CLIENT_REGISTERED);
	1691	+
	1692	+out:
	1693	+ up_write(&clients_rwsem);
	1694	+ return ret;
	1695	+}
	1696	+
	1697	+static void remove_client_id(struct ib_client *client)
	1698	+{
	1699	+ down_write(&clients_rwsem);
	1700	+ xa_erase(&clients, client->client_id);
	1701	+ for (; highest_client_id; highest_client_id--)
	1702	+ if (xa_load(&clients, highest_client_id - 1))
	1703	+ break;
	1704	+ up_write(&clients_rwsem);
	1705	+}
622	1706
623	1707	/**
624	1708	* ib_register_client - Register an IB client
..	..	@@ -636,19 +1720,25 @@
636	1720	int ib_register_client(struct ib_client *client)
637	1721	{
638	1722	struct ib_device *device;
	1723	+ unsigned long index;
	1724	+ int ret;
639	1725
640		- mutex_lock(&device_mutex);
	1726	+ refcount_set(&client->uses, 1);
	1727	+ init_completion(&client->uses_zero);
	1728	+ ret = assign_client_id(client);
	1729	+ if (ret)
	1730	+ return ret;
641	1731
642		- list_for_each_entry(device, &device_list, core_list)
643		- if (!add_client_context(device, client) && client->add)
644		- client->add(device);
645		-
646		- down_write(&lists_rwsem);
647		- list_add_tail(&client->list, &client_list);
648		- up_write(&lists_rwsem);
649		-
650		- mutex_unlock(&device_mutex);
651		-
	1732	+ down_read(&devices_rwsem);
	1733	+ xa_for_each_marked (&devices, index, device, DEVICE_REGISTERED) {
	1734	+ ret = add_client_context(device, client);
	1735	+ if (ret) {
	1736	+ up_read(&devices_rwsem);
	1737	+ ib_unregister_client(client);
	1738	+ return ret;
	1739	+ }
	1740	+ }
	1741	+ up_read(&devices_rwsem);
652	1742	return 0;
653	1743	}
654	1744	EXPORT_SYMBOL(ib_register_client);
..	..	@@ -660,80 +1750,140 @@
660	1750	* Upper level users use ib_unregister_client() to remove their client
661	1751	* registration. When ib_unregister_client() is called, the client
662	1752	* will receive a remove callback for each IB device still registered.
	1753	+ *
	1754	+ * This is a full fence, once it returns no client callbacks will be called,
	1755	+ * or are running in another thread.
663	1756	*/
664	1757	void ib_unregister_client(struct ib_client *client)
665	1758	{
666		- struct ib_client_data context, tmp;
667	1759	struct ib_device *device;
668		- unsigned long flags;
	1760	+ unsigned long index;
669	1761
670		- mutex_lock(&device_mutex);
	1762	+ down_write(&clients_rwsem);
	1763	+ ib_client_put(client);
	1764	+ xa_clear_mark(&clients, client->client_id, CLIENT_REGISTERED);
	1765	+ up_write(&clients_rwsem);
671	1766
672		- down_write(&lists_rwsem);
673		- list_del(&client->list);
674		- up_write(&lists_rwsem);
675		-
676		- list_for_each_entry(device, &device_list, core_list) {
677		- struct ib_client_data *found_context = NULL;
678		-
679		- down_write(&lists_rwsem);
680		- spin_lock_irqsave(&device->client_data_lock, flags);
681		- list_for_each_entry_safe(context, tmp, &device->client_data_list, list)
682		- if (context->client == client) {
683		- context->going_down = true;
684		- found_context = context;
685		- break;
686		- }
687		- spin_unlock_irqrestore(&device->client_data_lock, flags);
688		- up_write(&lists_rwsem);
689		-
690		- if (client->remove)
691		- client->remove(device, found_context ?
692		- found_context->data : NULL);
693		-
694		- if (!found_context) {
695		- pr_warn("No client context found for %s/%s\n",
696		- device->name, client->name);
	1767	+ /* We do not want to have locks while calling client->remove() */
	1768	+ rcu_read_lock();
	1769	+ xa_for_each (&devices, index, device) {
	1770	+ if (!ib_device_try_get(device))
697	1771	continue;
698		- }
	1772	+ rcu_read_unlock();
699	1773
700		- down_write(&lists_rwsem);
701		- spin_lock_irqsave(&device->client_data_lock, flags);
702		- list_del(&found_context->list);
703		- kfree(found_context);
704		- spin_unlock_irqrestore(&device->client_data_lock, flags);
705		- up_write(&lists_rwsem);
	1774	+ remove_client_context(device, client->client_id);
	1775	+
	1776	+ ib_device_put(device);
	1777	+ rcu_read_lock();
706	1778	}
	1779	+ rcu_read_unlock();
707	1780
708		- mutex_unlock(&device_mutex);
	1781	+ /*
	1782	+ * remove_client_context() is not a fence, it can return even though a
	1783	+ * removal is ongoing. Wait until all removals are completed.
	1784	+ */
	1785	+ wait_for_completion(&client->uses_zero);
	1786	+ remove_client_id(client);
709	1787	}
710	1788	EXPORT_SYMBOL(ib_unregister_client);
711	1789
712		-/**
713		- * ib_get_client_data - Get IB client context
714		- * @device:Device to get context for
715		- * @client:Client to get context for
716		- *
717		- * ib_get_client_data() returns client context set with
718		- * ib_set_client_data().
719		- */
720		-void ib_get_client_data(struct ib_device device, struct ib_client *client)
	1790	+static int __ib_get_global_client_nl_info(const char *client_name,
	1791	+ struct ib_client_nl_info *res)
721	1792	{
722		- struct ib_client_data *context;
723		- void *ret = NULL;
724		- unsigned long flags;
	1793	+ struct ib_client *client;
	1794	+ unsigned long index;
	1795	+ int ret = -ENOENT;
725	1796
726		- spin_lock_irqsave(&device->client_data_lock, flags);
727		- list_for_each_entry(context, &device->client_data_list, list)
728		- if (context->client == client) {
729		- ret = context->data;
	1797	+ down_read(&clients_rwsem);
	1798	+ xa_for_each_marked (&clients, index, client, CLIENT_REGISTERED) {
	1799	+ if (strcmp(client->name, client_name) != 0)
	1800	+ continue;
	1801	+ if (!client->get_global_nl_info) {
	1802	+ ret = -EOPNOTSUPP;
730	1803	break;
731	1804	}
732		- spin_unlock_irqrestore(&device->client_data_lock, flags);
	1805	+ ret = client->get_global_nl_info(res);
	1806	+ if (WARN_ON(ret == -ENOENT))
	1807	+ ret = -EINVAL;
	1808	+ if (!ret && res->cdev)
	1809	+ get_device(res->cdev);
	1810	+ break;
	1811	+ }
	1812	+ up_read(&clients_rwsem);
	1813	+ return ret;
	1814	+}
	1815	+
	1816	+static int __ib_get_client_nl_info(struct ib_device *ibdev,
	1817	+ const char *client_name,
	1818	+ struct ib_client_nl_info *res)
	1819	+{
	1820	+ unsigned long index;
	1821	+ void *client_data;
	1822	+ int ret = -ENOENT;
	1823	+
	1824	+ down_read(&ibdev->client_data_rwsem);
	1825	+ xan_for_each_marked (&ibdev->client_data, index, client_data,
	1826	+ CLIENT_DATA_REGISTERED) {
	1827	+ struct ib_client *client = xa_load(&clients, index);
	1828	+
	1829	+ if (!client \|\| strcmp(client->name, client_name) != 0)
	1830	+ continue;
	1831	+ if (!client->get_nl_info) {
	1832	+ ret = -EOPNOTSUPP;
	1833	+ break;
	1834	+ }
	1835	+ ret = client->get_nl_info(ibdev, client_data, res);
	1836	+ if (WARN_ON(ret == -ENOENT))
	1837	+ ret = -EINVAL;
	1838	+
	1839	+ /*
	1840	+ * The cdev is guaranteed valid as long as we are inside the
	1841	+ * client_data_rwsem as remove_one can't be called. Keep it
	1842	+ * valid for the caller.
	1843	+ */
	1844	+ if (!ret && res->cdev)
	1845	+ get_device(res->cdev);
	1846	+ break;
	1847	+ }
	1848	+ up_read(&ibdev->client_data_rwsem);
733	1849
734	1850	return ret;
735	1851	}
736		-EXPORT_SYMBOL(ib_get_client_data);
	1852	+
	1853	+/**
	1854	+ * ib_get_client_nl_info - Fetch the nl_info from a client
	1855	+ * @device - IB device
	1856	+ * @client_name - Name of the client
	1857	+ * @res - Result of the query
	1858	+ */
	1859	+int ib_get_client_nl_info(struct ib_device ibdev, const char client_name,
	1860	+ struct ib_client_nl_info *res)
	1861	+{
	1862	+ int ret;
	1863	+
	1864	+ if (ibdev)
	1865	+ ret = __ib_get_client_nl_info(ibdev, client_name, res);
	1866	+ else
	1867	+ ret = __ib_get_global_client_nl_info(client_name, res);
	1868	+#ifdef CONFIG_MODULES
	1869	+ if (ret == -ENOENT) {
	1870	+ request_module("rdma-client-%s", client_name);
	1871	+ if (ibdev)
	1872	+ ret = __ib_get_client_nl_info(ibdev, client_name, res);
	1873	+ else
	1874	+ ret = __ib_get_global_client_nl_info(client_name, res);
	1875	+ }
	1876	+#endif
	1877	+ if (ret) {
	1878	+ if (ret == -ENOENT)
	1879	+ return -EOPNOTSUPP;
	1880	+ return ret;
	1881	+ }
	1882	+
	1883	+ if (WARN_ON(!res->cdev))
	1884	+ return -EINVAL;
	1885	+ return 0;
	1886	+}
737	1887
738	1888	/**
739	1889	* ib_set_client_data - Set IB client context
..	..	@@ -741,27 +1891,22 @@
741	1891	* @client:Client to set context for
742	1892	* @data:Context to set
743	1893	*
744		- * ib_set_client_data() sets client context that can be retrieved with
745		- * ib_get_client_data().
	1894	+ * ib_set_client_data() sets client context data that can be retrieved with
	1895	+ * ib_get_client_data(). This can only be called while the client is
	1896	+ * registered to the device, once the ib_client remove() callback returns this
	1897	+ * cannot be called.
746	1898	*/
747	1899	void ib_set_client_data(struct ib_device device, struct ib_client client,
748	1900	void *data)
749	1901	{
750		- struct ib_client_data *context;
751		- unsigned long flags;
	1902	+ void *rc;
752	1903
753		- spin_lock_irqsave(&device->client_data_lock, flags);
754		- list_for_each_entry(context, &device->client_data_list, list)
755		- if (context->client == client) {
756		- context->data = data;
757		- goto out;
758		- }
	1904	+ if (WARN_ON(IS_ERR(data)))
	1905	+ data = NULL;
759	1906
760		- pr_warn("No client context found for %s/%s\n",
761		- device->name, client->name);
762		-
763		-out:
764		- spin_unlock_irqrestore(&device->client_data_lock, flags);
	1907	+ rc = xa_store(&device->client_data, client->client_id, data,
	1908	+ GFP_KERNEL);
	1909	+ WARN_ON(xa_is_err(rc));
765	1910	}
766	1911	EXPORT_SYMBOL(ib_set_client_data);
767	1912
..	..	@@ -771,17 +1916,15 @@
771	1916	*
772	1917	* ib_register_event_handler() registers an event handler that will be
773	1918	* called back when asynchronous IB events occur (as defined in
774		- * chapter 11 of the InfiniBand Architecture Specification). This
775		- * callback may occur in interrupt context.
	1919	+ * chapter 11 of the InfiniBand Architecture Specification). This
	1920	+ * callback occurs in workqueue context.
776	1921	*/
777	1922	void ib_register_event_handler(struct ib_event_handler *event_handler)
778	1923	{
779		- unsigned long flags;
780		-
781		- spin_lock_irqsave(&event_handler->device->event_handler_lock, flags);
	1924	+ down_write(&event_handler->device->event_handler_rwsem);
782	1925	list_add_tail(&event_handler->list,
783	1926	&event_handler->device->event_handler_list);
784		- spin_unlock_irqrestore(&event_handler->device->event_handler_lock, flags);
	1927	+ up_write(&event_handler->device->event_handler_rwsem);
785	1928	}
786	1929	EXPORT_SYMBOL(ib_register_event_handler);
787	1930
..	..	@@ -794,35 +1937,87 @@
794	1937	*/
795	1938	void ib_unregister_event_handler(struct ib_event_handler *event_handler)
796	1939	{
797		- unsigned long flags;
798		-
799		- spin_lock_irqsave(&event_handler->device->event_handler_lock, flags);
	1940	+ down_write(&event_handler->device->event_handler_rwsem);
800	1941	list_del(&event_handler->list);
801		- spin_unlock_irqrestore(&event_handler->device->event_handler_lock, flags);
	1942	+ up_write(&event_handler->device->event_handler_rwsem);
802	1943	}
803	1944	EXPORT_SYMBOL(ib_unregister_event_handler);
804	1945
805		-/**
806		- * ib_dispatch_event - Dispatch an asynchronous event
807		- * @event:Event to dispatch
808		- *
809		- * Low-level drivers must call ib_dispatch_event() to dispatch the
810		- * event to all registered event handlers when an asynchronous event
811		- * occurs.
812		- */
813		-void ib_dispatch_event(struct ib_event *event)
	1946	+void ib_dispatch_event_clients(struct ib_event *event)
814	1947	{
815		- unsigned long flags;
816	1948	struct ib_event_handler *handler;
817	1949
818		- spin_lock_irqsave(&event->device->event_handler_lock, flags);
	1950	+ down_read(&event->device->event_handler_rwsem);
819	1951
820	1952	list_for_each_entry(handler, &event->device->event_handler_list, list)
821	1953	handler->handler(handler, event);
822	1954
823		- spin_unlock_irqrestore(&event->device->event_handler_lock, flags);
	1955	+ up_read(&event->device->event_handler_rwsem);
824	1956	}
825		-EXPORT_SYMBOL(ib_dispatch_event);
	1957	+
	1958	+static int iw_query_port(struct ib_device *device,
	1959	+ u8 port_num,
	1960	+ struct ib_port_attr *port_attr)
	1961	+{
	1962	+ struct in_device *inetdev;
	1963	+ struct net_device *netdev;
	1964	+
	1965	+ memset(port_attr, 0, sizeof(*port_attr));
	1966	+
	1967	+ netdev = ib_device_get_netdev(device, port_num);
	1968	+ if (!netdev)
	1969	+ return -ENODEV;
	1970	+
	1971	+ port_attr->max_mtu = IB_MTU_4096;
	1972	+ port_attr->active_mtu = ib_mtu_int_to_enum(netdev->mtu);
	1973	+
	1974	+ if (!netif_carrier_ok(netdev)) {
	1975	+ port_attr->state = IB_PORT_DOWN;
	1976	+ port_attr->phys_state = IB_PORT_PHYS_STATE_DISABLED;
	1977	+ } else {
	1978	+ rcu_read_lock();
	1979	+ inetdev = __in_dev_get_rcu(netdev);
	1980	+
	1981	+ if (inetdev && inetdev->ifa_list) {
	1982	+ port_attr->state = IB_PORT_ACTIVE;
	1983	+ port_attr->phys_state = IB_PORT_PHYS_STATE_LINK_UP;
	1984	+ } else {
	1985	+ port_attr->state = IB_PORT_INIT;
	1986	+ port_attr->phys_state =
	1987	+ IB_PORT_PHYS_STATE_PORT_CONFIGURATION_TRAINING;
	1988	+ }
	1989	+
	1990	+ rcu_read_unlock();
	1991	+ }
	1992	+
	1993	+ dev_put(netdev);
	1994	+ return device->ops.query_port(device, port_num, port_attr);
	1995	+}
	1996	+
	1997	+static int __ib_query_port(struct ib_device *device,
	1998	+ u8 port_num,
	1999	+ struct ib_port_attr *port_attr)
	2000	+{
	2001	+ union ib_gid gid = {};
	2002	+ int err;
	2003	+
	2004	+ memset(port_attr, 0, sizeof(*port_attr));
	2005	+
	2006	+ err = device->ops.query_port(device, port_num, port_attr);
	2007	+ if (err \|\| port_attr->subnet_prefix)
	2008	+ return err;
	2009	+
	2010	+ if (rdma_port_get_link_layer(device, port_num) !=
	2011	+ IB_LINK_LAYER_INFINIBAND)
	2012	+ return 0;
	2013	+
	2014	+ err = device->ops.query_gid(device, port_num, 0, &gid);
	2015	+ if (err)
	2016	+ return err;
	2017	+
	2018	+ port_attr->subnet_prefix = be64_to_cpu(gid.global.subnet_prefix);
	2019	+ return 0;
	2020	+}
826	2021
827	2022	/**
828	2023	* ib_query_port - Query IB port attributes
..	..	@@ -837,28 +2032,197 @@
837	2032	u8 port_num,
838	2033	struct ib_port_attr *port_attr)
839	2034	{
840		- union ib_gid gid;
841		- int err;
842		-
843	2035	if (!rdma_is_port_valid(device, port_num))
844	2036	return -EINVAL;
845	2037
846		- memset(port_attr, 0, sizeof(*port_attr));
847		- err = device->query_port(device, port_num, port_attr);
848		- if (err \|\| port_attr->subnet_prefix)
849		- return err;
850		-
851		- if (rdma_port_get_link_layer(device, port_num) != IB_LINK_LAYER_INFINIBAND)
852		- return 0;
853		-
854		- err = device->query_gid(device, port_num, 0, &gid);
855		- if (err)
856		- return err;
857		-
858		- port_attr->subnet_prefix = be64_to_cpu(gid.global.subnet_prefix);
859		- return 0;
	2038	+ if (rdma_protocol_iwarp(device, port_num))
	2039	+ return iw_query_port(device, port_num, port_attr);
	2040	+ else
	2041	+ return __ib_query_port(device, port_num, port_attr);
860	2042	}
861	2043	EXPORT_SYMBOL(ib_query_port);
	2044	+
	2045	+static void add_ndev_hash(struct ib_port_data *pdata)
	2046	+{
	2047	+ unsigned long flags;
	2048	+
	2049	+ might_sleep();
	2050	+
	2051	+ spin_lock_irqsave(&ndev_hash_lock, flags);
	2052	+ if (hash_hashed(&pdata->ndev_hash_link)) {
	2053	+ hash_del_rcu(&pdata->ndev_hash_link);
	2054	+ spin_unlock_irqrestore(&ndev_hash_lock, flags);
	2055	+ /*
	2056	+ * We cannot do hash_add_rcu after a hash_del_rcu until the
	2057	+ * grace period
	2058	+ */
	2059	+ synchronize_rcu();
	2060	+ spin_lock_irqsave(&ndev_hash_lock, flags);
	2061	+ }
	2062	+ if (pdata->netdev)
	2063	+ hash_add_rcu(ndev_hash, &pdata->ndev_hash_link,
	2064	+ (uintptr_t)pdata->netdev);
	2065	+ spin_unlock_irqrestore(&ndev_hash_lock, flags);
	2066	+}
	2067	+
	2068	+/**
	2069	+ * ib_device_set_netdev - Associate the ib_dev with an underlying net_device
	2070	+ * @ib_dev: Device to modify
	2071	+ * @ndev: net_device to affiliate, may be NULL
	2072	+ * @port: IB port the net_device is connected to
	2073	+ *
	2074	+ * Drivers should use this to link the ib_device to a netdev so the netdev
	2075	+ * shows up in interfaces like ib_enum_roce_netdev. Only one netdev may be
	2076	+ * affiliated with any port.
	2077	+ *
	2078	+ * The caller must ensure that the given ndev is not unregistered or
	2079	+ * unregistering, and that either the ib_device is unregistered or
	2080	+ * ib_device_set_netdev() is called with NULL when the ndev sends a
	2081	+ * NETDEV_UNREGISTER event.
	2082	+ */
	2083	+int ib_device_set_netdev(struct ib_device ib_dev, struct net_device ndev,
	2084	+ unsigned int port)
	2085	+{
	2086	+ struct net_device *old_ndev;
	2087	+ struct ib_port_data *pdata;
	2088	+ unsigned long flags;
	2089	+ int ret;
	2090	+
	2091	+ /*
	2092	+ * Drivers wish to call this before ib_register_driver, so we have to
	2093	+ * setup the port data early.
	2094	+ */
	2095	+ ret = alloc_port_data(ib_dev);
	2096	+ if (ret)
	2097	+ return ret;
	2098	+
	2099	+ if (!rdma_is_port_valid(ib_dev, port))
	2100	+ return -EINVAL;
	2101	+
	2102	+ pdata = &ib_dev->port_data[port];
	2103	+ spin_lock_irqsave(&pdata->netdev_lock, flags);
	2104	+ old_ndev = rcu_dereference_protected(
	2105	+ pdata->netdev, lockdep_is_held(&pdata->netdev_lock));
	2106	+ if (old_ndev == ndev) {
	2107	+ spin_unlock_irqrestore(&pdata->netdev_lock, flags);
	2108	+ return 0;
	2109	+ }
	2110	+
	2111	+ if (ndev)
	2112	+ dev_hold(ndev);
	2113	+ rcu_assign_pointer(pdata->netdev, ndev);
	2114	+ spin_unlock_irqrestore(&pdata->netdev_lock, flags);
	2115	+
	2116	+ add_ndev_hash(pdata);
	2117	+ if (old_ndev)
	2118	+ dev_put(old_ndev);
	2119	+
	2120	+ return 0;
	2121	+}
	2122	+EXPORT_SYMBOL(ib_device_set_netdev);
	2123	+
	2124	+static void free_netdevs(struct ib_device *ib_dev)
	2125	+{
	2126	+ unsigned long flags;
	2127	+ unsigned int port;
	2128	+
	2129	+ if (!ib_dev->port_data)
	2130	+ return;
	2131	+
	2132	+ rdma_for_each_port (ib_dev, port) {
	2133	+ struct ib_port_data *pdata = &ib_dev->port_data[port];
	2134	+ struct net_device *ndev;
	2135	+
	2136	+ spin_lock_irqsave(&pdata->netdev_lock, flags);
	2137	+ ndev = rcu_dereference_protected(
	2138	+ pdata->netdev, lockdep_is_held(&pdata->netdev_lock));
	2139	+ if (ndev) {
	2140	+ spin_lock(&ndev_hash_lock);
	2141	+ hash_del_rcu(&pdata->ndev_hash_link);
	2142	+ spin_unlock(&ndev_hash_lock);
	2143	+
	2144	+ /*
	2145	+ * If this is the last dev_put there is still a
	2146	+ * synchronize_rcu before the netdev is kfreed, so we
	2147	+ * can continue to rely on unlocked pointer
	2148	+ * comparisons after the put
	2149	+ */
	2150	+ rcu_assign_pointer(pdata->netdev, NULL);
	2151	+ dev_put(ndev);
	2152	+ }
	2153	+ spin_unlock_irqrestore(&pdata->netdev_lock, flags);
	2154	+ }
	2155	+}
	2156	+
	2157	+struct net_device ib_device_get_netdev(struct ib_device ib_dev,
	2158	+ unsigned int port)
	2159	+{
	2160	+ struct ib_port_data *pdata;
	2161	+ struct net_device *res;
	2162	+
	2163	+ if (!rdma_is_port_valid(ib_dev, port))
	2164	+ return NULL;
	2165	+
	2166	+ pdata = &ib_dev->port_data[port];
	2167	+
	2168	+ /*
	2169	+ * New drivers should use ib_device_set_netdev() not the legacy
	2170	+ * get_netdev().
	2171	+ */
	2172	+ if (ib_dev->ops.get_netdev)
	2173	+ res = ib_dev->ops.get_netdev(ib_dev, port);
	2174	+ else {
	2175	+ spin_lock(&pdata->netdev_lock);
	2176	+ res = rcu_dereference_protected(
	2177	+ pdata->netdev, lockdep_is_held(&pdata->netdev_lock));
	2178	+ if (res)
	2179	+ dev_hold(res);
	2180	+ spin_unlock(&pdata->netdev_lock);
	2181	+ }
	2182	+
	2183	+ /*
	2184	+ * If we are starting to unregister expedite things by preventing
	2185	+ * propagation of an unregistering netdev.
	2186	+ */
	2187	+ if (res && res->reg_state != NETREG_REGISTERED) {
	2188	+ dev_put(res);
	2189	+ return NULL;
	2190	+ }
	2191	+
	2192	+ return res;
	2193	+}
	2194	+
	2195	+/**
	2196	+ * ib_device_get_by_netdev - Find an IB device associated with a netdev
	2197	+ * @ndev: netdev to locate
	2198	+ * @driver_id: The driver ID that must match (RDMA_DRIVER_UNKNOWN matches all)
	2199	+ *
	2200	+ * Find and hold an ib_device that is associated with a netdev via
	2201	+ * ib_device_set_netdev(). The caller must call ib_device_put() on the
	2202	+ * returned pointer.
	2203	+ */
	2204	+struct ib_device ib_device_get_by_netdev(struct net_device ndev,
	2205	+ enum rdma_driver_id driver_id)
	2206	+{
	2207	+ struct ib_device *res = NULL;
	2208	+ struct ib_port_data *cur;
	2209	+
	2210	+ rcu_read_lock();
	2211	+ hash_for_each_possible_rcu (ndev_hash, cur, ndev_hash_link,
	2212	+ (uintptr_t)ndev) {
	2213	+ if (rcu_access_pointer(cur->netdev) == ndev &&
	2214	+ (driver_id == RDMA_DRIVER_UNKNOWN \|\|
	2215	+ cur->ib_dev->ops.driver_id == driver_id) &&
	2216	+ ib_device_try_get(cur->ib_dev)) {
	2217	+ res = cur->ib_dev;
	2218	+ break;
	2219	+ }
	2220	+ }
	2221	+ rcu_read_unlock();
	2222	+
	2223	+ return res;
	2224	+}
	2225	+EXPORT_SYMBOL(ib_device_get_by_netdev);
862	2226
863	2227	/**
864	2228	* ib_enum_roce_netdev - enumerate all RoCE ports
..	..	@@ -878,21 +2242,12 @@
878	2242	roce_netdev_callback cb,
879	2243	void *cookie)
880	2244	{
881		- u8 port;
	2245	+ unsigned int port;
882	2246
883		- for (port = rdma_start_port(ib_dev); port <= rdma_end_port(ib_dev);
884		- port++)
	2247	+ rdma_for_each_port (ib_dev, port)
885	2248	if (rdma_protocol_roce(ib_dev, port)) {
886		- struct net_device *idev = NULL;
887		-
888		- if (ib_dev->get_netdev)
889		- idev = ib_dev->get_netdev(ib_dev, port);
890		-
891		- if (idev &&
892		- idev->reg_state >= NETREG_UNREGISTERED) {
893		- dev_put(idev);
894		- idev = NULL;
895		- }
	2249	+ struct net_device *idev =
	2250	+ ib_device_get_netdev(ib_dev, port);
896	2251
897	2252	if (filter(ib_dev, port, idev, filter_cookie))
898	2253	cb(ib_dev, port, idev, cookie);
..	..	@@ -919,11 +2274,12 @@
919	2274	void *cookie)
920	2275	{
921	2276	struct ib_device *dev;
	2277	+ unsigned long index;
922	2278
923		- down_read(&lists_rwsem);
924		- list_for_each_entry(dev, &device_list, core_list)
	2279	+ down_read(&devices_rwsem);
	2280	+ xa_for_each_marked (&devices, index, dev, DEVICE_REGISTERED)
925	2281	ib_enum_roce_netdev(dev, filter, filter_cookie, cb, cookie);
926		- up_read(&lists_rwsem);
	2282	+ up_read(&devices_rwsem);
927	2283	}
928	2284
929	2285	/**
..	..	@@ -935,19 +2291,22 @@
935	2291	int ib_enum_all_devs(nldev_callback nldev_cb, struct sk_buff *skb,
936	2292	struct netlink_callback *cb)
937	2293	{
	2294	+ unsigned long index;
938	2295	struct ib_device *dev;
939	2296	unsigned int idx = 0;
940	2297	int ret = 0;
941	2298
942		- down_read(&lists_rwsem);
943		- list_for_each_entry(dev, &device_list, core_list) {
	2299	+ down_read(&devices_rwsem);
	2300	+ xa_for_each_marked (&devices, index, dev, DEVICE_REGISTERED) {
	2301	+ if (!rdma_dev_access_netns(dev, sock_net(skb->sk)))
	2302	+ continue;
	2303	+
944	2304	ret = nldev_cb(dev, skb, cb, idx);
945	2305	if (ret)
946	2306	break;
947	2307	idx++;
948	2308	}
949		-
950		- up_read(&lists_rwsem);
	2309	+ up_read(&devices_rwsem);
951	2310	return ret;
952	2311	}
953	2312
..	..	@@ -963,7 +2322,13 @@
963	2322	int ib_query_pkey(struct ib_device *device,
964	2323	u8 port_num, u16 index, u16 *pkey)
965	2324	{
966		- return device->query_pkey(device, port_num, index, pkey);
	2325	+ if (!rdma_is_port_valid(device, port_num))
	2326	+ return -EINVAL;
	2327	+
	2328	+ if (!device->ops.query_pkey)
	2329	+ return -EOPNOTSUPP;
	2330	+
	2331	+ return device->ops.query_pkey(device, port_num, index, pkey);
967	2332	}
968	2333	EXPORT_SYMBOL(ib_query_pkey);
969	2334
..	..	@@ -980,11 +2345,11 @@
980	2345	int device_modify_mask,
981	2346	struct ib_device_modify *device_modify)
982	2347	{
983		- if (!device->modify_device)
984		- return -ENOSYS;
	2348	+ if (!device->ops.modify_device)
	2349	+ return -EOPNOTSUPP;
985	2350
986		- return device->modify_device(device, device_modify_mask,
987		- device_modify);
	2351	+ return device->ops.modify_device(device, device_modify_mask,
	2352	+ device_modify);
988	2353	}
989	2354	EXPORT_SYMBOL(ib_modify_device);
990	2355
..	..	@@ -1008,11 +2373,16 @@
1008	2373	if (!rdma_is_port_valid(device, port_num))
1009	2374	return -EINVAL;
1010	2375
1011		- if (device->modify_port)
1012		- rc = device->modify_port(device, port_num, port_modify_mask,
1013		- port_modify);
	2376	+ if (device->ops.modify_port)
	2377	+ rc = device->ops.modify_port(device, port_num,
	2378	+ port_modify_mask,
	2379	+ port_modify);
	2380	+ else if (rdma_protocol_roce(device, port_num) &&
	2381	+ ((port_modify->set_port_cap_mask & ~IB_PORT_CM_SUP) == 0 \|\|
	2382	+ (port_modify->clr_port_cap_mask & ~IB_PORT_CM_SUP) == 0))
	2383	+ rc = 0;
1014	2384	else
1015		- rc = rdma_protocol_roce(device, port_num) ? 0 : -ENOSYS;
	2385	+ rc = -EOPNOTSUPP;
1016	2386	return rc;
1017	2387	}
1018	2388	EXPORT_SYMBOL(ib_modify_port);
..	..	@@ -1030,13 +2400,15 @@
1030	2400	u8 port_num, u16 index)
1031	2401	{
1032	2402	union ib_gid tmp_gid;
1033		- int ret, port, i;
	2403	+ unsigned int port;
	2404	+ int ret, i;
1034	2405
1035		- for (port = rdma_start_port(device); port <= rdma_end_port(device); ++port) {
	2406	+ rdma_for_each_port (device, port) {
1036	2407	if (!rdma_protocol_ib(device, port))
1037	2408	continue;
1038	2409
1039		- for (i = 0; i < device->port_immutable[port].gid_tbl_len; ++i) {
	2410	+ for (i = 0; i < device->port_data[port].immutable.gid_tbl_len;
	2411	+ ++i) {
1040	2412	ret = rdma_query_gid(device, port, i, &tmp_gid);
1041	2413	if (ret)
1042	2414	continue;
..	..	@@ -1069,7 +2441,8 @@
1069	2441	u16 tmp_pkey;
1070	2442	int partial_ix = -1;
1071	2443
1072		- for (i = 0; i < device->port_immutable[port_num].pkey_tbl_len; ++i) {
	2444	+ for (i = 0; i < device->port_data[port_num].immutable.pkey_tbl_len;
	2445	+ ++i) {
1073	2446	ret = ib_query_pkey(device, port_num, i, &tmp_pkey);
1074	2447	if (ret)
1075	2448	return ret;
..	..	@@ -1102,6 +2475,7 @@
1102	2475	* @gid: A GID that the net_dev uses to communicate.
1103	2476	* @addr: Contains the IP address that the request specified as its
1104	2477	* destination.
	2478	+ *
1105	2479	*/
1106	2480	struct net_device ib_get_net_dev_by_params(struct ib_device dev,
1107	2481	u8 port,
..	..	@@ -1110,33 +2484,201 @@
1110	2484	const struct sockaddr *addr)
1111	2485	{
1112	2486	struct net_device *net_dev = NULL;
1113		- struct ib_client_data *context;
	2487	+ unsigned long index;
	2488	+ void *client_data;
1114	2489
1115	2490	if (!rdma_protocol_ib(dev, port))
1116	2491	return NULL;
1117	2492
1118		- down_read(&lists_rwsem);
	2493	+ /*
	2494	+ * Holding the read side guarantees that the client will not become
	2495	+ * unregistered while we are calling get_net_dev_by_params()
	2496	+ */
	2497	+ down_read(&dev->client_data_rwsem);
	2498	+ xan_for_each_marked (&dev->client_data, index, client_data,
	2499	+ CLIENT_DATA_REGISTERED) {
	2500	+ struct ib_client *client = xa_load(&clients, index);
1119	2501
1120		- list_for_each_entry(context, &dev->client_data_list, list) {
1121		- struct ib_client *client = context->client;
1122		-
1123		- if (context->going_down)
	2502	+ if (!client \|\| !client->get_net_dev_by_params)
1124	2503	continue;
1125	2504
1126		- if (client->get_net_dev_by_params) {
1127		- net_dev = client->get_net_dev_by_params(dev, port, pkey,
1128		- gid, addr,
1129		- context->data);
1130		- if (net_dev)
1131		- break;
1132		- }
	2505	+ net_dev = client->get_net_dev_by_params(dev, port, pkey, gid,
	2506	+ addr, client_data);
	2507	+ if (net_dev)
	2508	+ break;
1133	2509	}
1134		-
1135		- up_read(&lists_rwsem);
	2510	+ up_read(&dev->client_data_rwsem);
1136	2511
1137	2512	return net_dev;
1138	2513	}
1139	2514	EXPORT_SYMBOL(ib_get_net_dev_by_params);
	2515	+
	2516	+void ib_set_device_ops(struct ib_device dev, const struct ib_device_ops ops)
	2517	+{
	2518	+ struct ib_device_ops *dev_ops = &dev->ops;
	2519	+#define SET_DEVICE_OP(ptr, name) \
	2520	+ do { \
	2521	+ if (ops->name) \
	2522	+ if (!((ptr)->name)) \
	2523	+ (ptr)->name = ops->name; \
	2524	+ } while (0)
	2525	+
	2526	+#define SET_OBJ_SIZE(ptr, name) SET_DEVICE_OP(ptr, size_##name)
	2527	+
	2528	+ if (ops->driver_id != RDMA_DRIVER_UNKNOWN) {
	2529	+ WARN_ON(dev_ops->driver_id != RDMA_DRIVER_UNKNOWN &&
	2530	+ dev_ops->driver_id != ops->driver_id);
	2531	+ dev_ops->driver_id = ops->driver_id;
	2532	+ }
	2533	+ if (ops->owner) {
	2534	+ WARN_ON(dev_ops->owner && dev_ops->owner != ops->owner);
	2535	+ dev_ops->owner = ops->owner;
	2536	+ }
	2537	+ if (ops->uverbs_abi_ver)
	2538	+ dev_ops->uverbs_abi_ver = ops->uverbs_abi_ver;
	2539	+
	2540	+ dev_ops->uverbs_no_driver_id_binding \|=
	2541	+ ops->uverbs_no_driver_id_binding;
	2542	+
	2543	+ SET_DEVICE_OP(dev_ops, add_gid);
	2544	+ SET_DEVICE_OP(dev_ops, advise_mr);
	2545	+ SET_DEVICE_OP(dev_ops, alloc_dm);
	2546	+ SET_DEVICE_OP(dev_ops, alloc_hw_stats);
	2547	+ SET_DEVICE_OP(dev_ops, alloc_mr);
	2548	+ SET_DEVICE_OP(dev_ops, alloc_mr_integrity);
	2549	+ SET_DEVICE_OP(dev_ops, alloc_mw);
	2550	+ SET_DEVICE_OP(dev_ops, alloc_pd);
	2551	+ SET_DEVICE_OP(dev_ops, alloc_rdma_netdev);
	2552	+ SET_DEVICE_OP(dev_ops, alloc_ucontext);
	2553	+ SET_DEVICE_OP(dev_ops, alloc_xrcd);
	2554	+ SET_DEVICE_OP(dev_ops, attach_mcast);
	2555	+ SET_DEVICE_OP(dev_ops, check_mr_status);
	2556	+ SET_DEVICE_OP(dev_ops, counter_alloc_stats);
	2557	+ SET_DEVICE_OP(dev_ops, counter_bind_qp);
	2558	+ SET_DEVICE_OP(dev_ops, counter_dealloc);
	2559	+ SET_DEVICE_OP(dev_ops, counter_unbind_qp);
	2560	+ SET_DEVICE_OP(dev_ops, counter_update_stats);
	2561	+ SET_DEVICE_OP(dev_ops, create_ah);
	2562	+ SET_DEVICE_OP(dev_ops, create_counters);
	2563	+ SET_DEVICE_OP(dev_ops, create_cq);
	2564	+ SET_DEVICE_OP(dev_ops, create_flow);
	2565	+ SET_DEVICE_OP(dev_ops, create_flow_action_esp);
	2566	+ SET_DEVICE_OP(dev_ops, create_qp);
	2567	+ SET_DEVICE_OP(dev_ops, create_rwq_ind_table);
	2568	+ SET_DEVICE_OP(dev_ops, create_srq);
	2569	+ SET_DEVICE_OP(dev_ops, create_wq);
	2570	+ SET_DEVICE_OP(dev_ops, dealloc_dm);
	2571	+ SET_DEVICE_OP(dev_ops, dealloc_driver);
	2572	+ SET_DEVICE_OP(dev_ops, dealloc_mw);
	2573	+ SET_DEVICE_OP(dev_ops, dealloc_pd);
	2574	+ SET_DEVICE_OP(dev_ops, dealloc_ucontext);
	2575	+ SET_DEVICE_OP(dev_ops, dealloc_xrcd);
	2576	+ SET_DEVICE_OP(dev_ops, del_gid);
	2577	+ SET_DEVICE_OP(dev_ops, dereg_mr);
	2578	+ SET_DEVICE_OP(dev_ops, destroy_ah);
	2579	+ SET_DEVICE_OP(dev_ops, destroy_counters);
	2580	+ SET_DEVICE_OP(dev_ops, destroy_cq);
	2581	+ SET_DEVICE_OP(dev_ops, destroy_flow);
	2582	+ SET_DEVICE_OP(dev_ops, destroy_flow_action);
	2583	+ SET_DEVICE_OP(dev_ops, destroy_qp);
	2584	+ SET_DEVICE_OP(dev_ops, destroy_rwq_ind_table);
	2585	+ SET_DEVICE_OP(dev_ops, destroy_srq);
	2586	+ SET_DEVICE_OP(dev_ops, destroy_wq);
	2587	+ SET_DEVICE_OP(dev_ops, detach_mcast);
	2588	+ SET_DEVICE_OP(dev_ops, disassociate_ucontext);
	2589	+ SET_DEVICE_OP(dev_ops, drain_rq);
	2590	+ SET_DEVICE_OP(dev_ops, drain_sq);
	2591	+ SET_DEVICE_OP(dev_ops, enable_driver);
	2592	+ SET_DEVICE_OP(dev_ops, fill_res_cm_id_entry);
	2593	+ SET_DEVICE_OP(dev_ops, fill_res_cq_entry);
	2594	+ SET_DEVICE_OP(dev_ops, fill_res_cq_entry_raw);
	2595	+ SET_DEVICE_OP(dev_ops, fill_res_mr_entry);
	2596	+ SET_DEVICE_OP(dev_ops, fill_res_mr_entry_raw);
	2597	+ SET_DEVICE_OP(dev_ops, fill_res_qp_entry);
	2598	+ SET_DEVICE_OP(dev_ops, fill_res_qp_entry_raw);
	2599	+ SET_DEVICE_OP(dev_ops, fill_stat_mr_entry);
	2600	+ SET_DEVICE_OP(dev_ops, get_dev_fw_str);
	2601	+ SET_DEVICE_OP(dev_ops, get_dma_mr);
	2602	+ SET_DEVICE_OP(dev_ops, get_hw_stats);
	2603	+ SET_DEVICE_OP(dev_ops, get_link_layer);
	2604	+ SET_DEVICE_OP(dev_ops, get_netdev);
	2605	+ SET_DEVICE_OP(dev_ops, get_port_immutable);
	2606	+ SET_DEVICE_OP(dev_ops, get_vector_affinity);
	2607	+ SET_DEVICE_OP(dev_ops, get_vf_config);
	2608	+ SET_DEVICE_OP(dev_ops, get_vf_guid);
	2609	+ SET_DEVICE_OP(dev_ops, get_vf_stats);
	2610	+ SET_DEVICE_OP(dev_ops, init_port);
	2611	+ SET_DEVICE_OP(dev_ops, iw_accept);
	2612	+ SET_DEVICE_OP(dev_ops, iw_add_ref);
	2613	+ SET_DEVICE_OP(dev_ops, iw_connect);
	2614	+ SET_DEVICE_OP(dev_ops, iw_create_listen);
	2615	+ SET_DEVICE_OP(dev_ops, iw_destroy_listen);
	2616	+ SET_DEVICE_OP(dev_ops, iw_get_qp);
	2617	+ SET_DEVICE_OP(dev_ops, iw_reject);
	2618	+ SET_DEVICE_OP(dev_ops, iw_rem_ref);
	2619	+ SET_DEVICE_OP(dev_ops, map_mr_sg);
	2620	+ SET_DEVICE_OP(dev_ops, map_mr_sg_pi);
	2621	+ SET_DEVICE_OP(dev_ops, mmap);
	2622	+ SET_DEVICE_OP(dev_ops, mmap_free);
	2623	+ SET_DEVICE_OP(dev_ops, modify_ah);
	2624	+ SET_DEVICE_OP(dev_ops, modify_cq);
	2625	+ SET_DEVICE_OP(dev_ops, modify_device);
	2626	+ SET_DEVICE_OP(dev_ops, modify_flow_action_esp);
	2627	+ SET_DEVICE_OP(dev_ops, modify_port);
	2628	+ SET_DEVICE_OP(dev_ops, modify_qp);
	2629	+ SET_DEVICE_OP(dev_ops, modify_srq);
	2630	+ SET_DEVICE_OP(dev_ops, modify_wq);
	2631	+ SET_DEVICE_OP(dev_ops, peek_cq);
	2632	+ SET_DEVICE_OP(dev_ops, poll_cq);
	2633	+ SET_DEVICE_OP(dev_ops, post_recv);
	2634	+ SET_DEVICE_OP(dev_ops, post_send);
	2635	+ SET_DEVICE_OP(dev_ops, post_srq_recv);
	2636	+ SET_DEVICE_OP(dev_ops, process_mad);
	2637	+ SET_DEVICE_OP(dev_ops, query_ah);
	2638	+ SET_DEVICE_OP(dev_ops, query_device);
	2639	+ SET_DEVICE_OP(dev_ops, query_gid);
	2640	+ SET_DEVICE_OP(dev_ops, query_pkey);
	2641	+ SET_DEVICE_OP(dev_ops, query_port);
	2642	+ SET_DEVICE_OP(dev_ops, query_qp);
	2643	+ SET_DEVICE_OP(dev_ops, query_srq);
	2644	+ SET_DEVICE_OP(dev_ops, query_ucontext);
	2645	+ SET_DEVICE_OP(dev_ops, rdma_netdev_get_params);
	2646	+ SET_DEVICE_OP(dev_ops, read_counters);
	2647	+ SET_DEVICE_OP(dev_ops, reg_dm_mr);
	2648	+ SET_DEVICE_OP(dev_ops, reg_user_mr);
	2649	+ SET_DEVICE_OP(dev_ops, req_ncomp_notif);
	2650	+ SET_DEVICE_OP(dev_ops, req_notify_cq);
	2651	+ SET_DEVICE_OP(dev_ops, rereg_user_mr);
	2652	+ SET_DEVICE_OP(dev_ops, resize_cq);
	2653	+ SET_DEVICE_OP(dev_ops, set_vf_guid);
	2654	+ SET_DEVICE_OP(dev_ops, set_vf_link_state);
	2655	+
	2656	+ SET_OBJ_SIZE(dev_ops, ib_ah);
	2657	+ SET_OBJ_SIZE(dev_ops, ib_counters);
	2658	+ SET_OBJ_SIZE(dev_ops, ib_cq);
	2659	+ SET_OBJ_SIZE(dev_ops, ib_mw);
	2660	+ SET_OBJ_SIZE(dev_ops, ib_pd);
	2661	+ SET_OBJ_SIZE(dev_ops, ib_rwq_ind_table);
	2662	+ SET_OBJ_SIZE(dev_ops, ib_srq);
	2663	+ SET_OBJ_SIZE(dev_ops, ib_ucontext);
	2664	+ SET_OBJ_SIZE(dev_ops, ib_xrcd);
	2665	+}
	2666	+EXPORT_SYMBOL(ib_set_device_ops);
	2667	+
	2668	+#ifdef CONFIG_INFINIBAND_VIRT_DMA
	2669	+int ib_dma_virt_map_sg(struct ib_device dev, struct scatterlist sg, int nents)
	2670	+{
	2671	+ struct scatterlist *s;
	2672	+ int i;
	2673	+
	2674	+ for_each_sg(sg, s, nents, i) {
	2675	+ sg_dma_address(s) = (uintptr_t)sg_virt(s);
	2676	+ sg_dma_len(s) = s->length;
	2677	+ }
	2678	+ return nents;
	2679	+}
	2680	+EXPORT_SYMBOL(ib_dma_virt_map_sg);
	2681	+#endif /* CONFIG_INFINIBAND_VIRT_DMA */
1140	2682
1141	2683	static const struct rdma_nl_cbs ibnl_ls_cb_table[RDMA_NL_LS_NUM_OPS] = {
1142	2684	[RDMA_NL_LS_OP_RESOLVE] = {
..	..	@@ -1183,15 +2725,11 @@
1183	2725	goto err_comp_unbound;
1184	2726	}
1185	2727
1186		- ret = rdma_nl_init();
1187		- if (ret) {
1188		- pr_warn("Couldn't init IB netlink interface: err %d\n", ret);
1189		- goto err_sysfs;
1190		- }
	2728	+ rdma_nl_init();
1191	2729
1192	2730	ret = addr_init();
1193	2731	if (ret) {
1194		- pr_warn("Could't init IB address resolution\n");
	2732	+ pr_warn("Couldn't init IB address resolution\n");
1195	2733	goto err_ibnl;
1196	2734	}
1197	2735
..	..	@@ -1207,18 +2745,34 @@
1207	2745	goto err_mad;
1208	2746	}
1209	2747
1210		- ret = register_lsm_notifier(&ibdev_lsm_nb);
	2748	+ ret = register_blocking_lsm_notifier(&ibdev_lsm_nb);
1211	2749	if (ret) {
1212	2750	pr_warn("Couldn't register LSM notifier. ret %d\n", ret);
1213	2751	goto err_sa;
1214	2752	}
1215	2753
	2754	+ ret = register_pernet_device(&rdma_dev_net_ops);
	2755	+ if (ret) {
	2756	+ pr_warn("Couldn't init compat dev. ret %d\n", ret);
	2757	+ goto err_compat;
	2758	+ }
	2759	+
1216	2760	nldev_init();
1217	2761	rdma_nl_register(RDMA_NL_LS, ibnl_ls_cb_table);
1218		- roce_gid_mgmt_init();
	2762	+ ret = roce_gid_mgmt_init();
	2763	+ if (ret) {
	2764	+ pr_warn("Couldn't init RoCE GID management\n");
	2765	+ goto err_parent;
	2766	+ }
1219	2767
1220	2768	return 0;
1221	2769
	2770	+err_parent:
	2771	+ rdma_nl_unregister(RDMA_NL_LS);
	2772	+ nldev_exit();
	2773	+ unregister_pernet_device(&rdma_dev_net_ops);
	2774	+err_compat:
	2775	+ unregister_blocking_lsm_notifier(&ibdev_lsm_nb);
1222	2776	err_sa:
1223	2777	ib_sa_cleanup();
1224	2778	err_mad:
..	..	@@ -1226,8 +2780,6 @@
1226	2780	err_addr:
1227	2781	addr_cleanup();
1228	2782	err_ibnl:
1229		- rdma_nl_exit();
1230		-err_sysfs:
1231	2783	class_unregister(&ib_class);
1232	2784	err_comp_unbound:
1233	2785	destroy_workqueue(ib_comp_unbound_wq);
..	..	@@ -1243,7 +2795,8 @@
1243	2795	roce_gid_mgmt_cleanup();
1244	2796	nldev_exit();
1245	2797	rdma_nl_unregister(RDMA_NL_LS);
1246		- unregister_lsm_notifier(&ibdev_lsm_nb);
	2798	+ unregister_pernet_device(&rdma_dev_net_ops);
	2799	+ unregister_blocking_lsm_notifier(&ibdev_lsm_nb);
1247	2800	ib_sa_cleanup();
1248	2801	ib_mad_cleanup();
1249	2802	addr_cleanup();
..	..	@@ -1253,9 +2806,15 @@
1253	2806	destroy_workqueue(ib_comp_wq);
1254	2807	/* Make sure that any pending umem accounting work is done. */
1255	2808	destroy_workqueue(ib_wq);
	2809	+ flush_workqueue(system_unbound_wq);
	2810	+ WARN_ON(!xa_empty(&clients));
	2811	+ WARN_ON(!xa_empty(&devices));
1256	2812	}
1257	2813
1258	2814	MODULE_ALIAS_RDMA_NETLINK(RDMA_NL_LS, 4);
1259	2815
1260		-subsys_initcall(ib_core_init);
	2816	+/* ib core relies on netdev stack to first register net_ns_type_operations
	2817	+ * ns kobject type before ib_core initialization.
	2818	+ */
	2819	+fs_initcall(ib_core_init);
1261	2820	module_exit(ib_core_cleanup);