~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 +577,43 @@
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);
	602	+
	603	+ device->uverbs_ex_cmd_mask =
	604	+ BIT_ULL(IB_USER_VERBS_EX_CMD_CREATE_FLOW) \|
	605	+ BIT_ULL(IB_USER_VERBS_EX_CMD_CREATE_RWQ_IND_TBL) \|
	606	+ BIT_ULL(IB_USER_VERBS_EX_CMD_CREATE_WQ) \|
	607	+ BIT_ULL(IB_USER_VERBS_EX_CMD_DESTROY_FLOW) \|
	608	+ BIT_ULL(IB_USER_VERBS_EX_CMD_DESTROY_RWQ_IND_TBL) \|
	609	+ BIT_ULL(IB_USER_VERBS_EX_CMD_DESTROY_WQ) \|
	610	+ BIT_ULL(IB_USER_VERBS_EX_CMD_MODIFY_CQ) \|
	611	+ BIT_ULL(IB_USER_VERBS_EX_CMD_MODIFY_WQ) \|
	612	+ BIT_ULL(IB_USER_VERBS_EX_CMD_QUERY_DEVICE);
276	613
277	614	return device;
278	615	}
279		-EXPORT_SYMBOL(ib_alloc_device);
	616	+EXPORT_SYMBOL(_ib_alloc_device);
280	617
281	618	/**
282	619	* ib_dealloc_device - free an IB device struct
..	..	@@ -286,32 +623,173 @@
286	623	*/
287	624	void ib_dealloc_device(struct ib_device *device)
288	625	{
289		- WARN_ON(device->reg_state != IB_DEV_UNREGISTERED &&
290		- device->reg_state != IB_DEV_UNINITIALIZED);
291		- rdma_restrack_clean(&device->res);
	626	+ if (device->ops.dealloc_driver)
	627	+ device->ops.dealloc_driver(device);
	628	+
	629	+ /*
	630	+ * ib_unregister_driver() requires all devices to remain in the xarray
	631	+ * while their ops are callable. The last op we call is dealloc_driver
	632	+ * above. This is needed to create a fence on op callbacks prior to
	633	+ * allowing the driver module to unload.
	634	+ */
	635	+ down_write(&devices_rwsem);
	636	+ if (xa_load(&devices, device->index) == device)
	637	+ xa_erase(&devices, device->index);
	638	+ up_write(&devices_rwsem);
	639	+
	640	+ /* Expedite releasing netdev references */
	641	+ free_netdevs(device);
	642	+
	643	+ WARN_ON(!xa_empty(&device->compat_devs));
	644	+ WARN_ON(!xa_empty(&device->client_data));
	645	+ WARN_ON(refcount_read(&device->refcount));
	646	+ rdma_restrack_clean(device);
	647	+ /* Balances with device_initialize */
292	648	put_device(&device->dev);
293	649	}
294	650	EXPORT_SYMBOL(ib_dealloc_device);
295	651
296		-static int add_client_context(struct ib_device device, struct ib_client client)
	652	+/*
	653	+ * add_client_context() and remove_client_context() must be safe against
	654	+ * parallel calls on the same device - registration/unregistration of both the
	655	+ * device and client can be occurring in parallel.
	656	+ *
	657	+ * The routines need to be a fence, any caller must not return until the add
	658	+ * or remove is fully completed.
	659	+ */
	660	+static int add_client_context(struct ib_device *device,
	661	+ struct ib_client *client)
297	662	{
298		- struct ib_client_data *context;
299		- unsigned long flags;
	663	+ int ret = 0;
300	664
301		- context = kmalloc(sizeof *context, GFP_KERNEL);
302		- if (!context)
	665	+ if (!device->kverbs_provider && !client->no_kverbs_req)
	666	+ return 0;
	667	+
	668	+ down_write(&device->client_data_rwsem);
	669	+ /*
	670	+ * So long as the client is registered hold both the client and device
	671	+ * unregistration locks.
	672	+ */
	673	+ if (!refcount_inc_not_zero(&client->uses))
	674	+ goto out_unlock;
	675	+ refcount_inc(&device->refcount);
	676	+
	677	+ /*
	678	+ * Another caller to add_client_context got here first and has already
	679	+ * completely initialized context.
	680	+ */
	681	+ if (xa_get_mark(&device->client_data, client->client_id,
	682	+ CLIENT_DATA_REGISTERED))
	683	+ goto out;
	684	+
	685	+ ret = xa_err(xa_store(&device->client_data, client->client_id, NULL,
	686	+ GFP_KERNEL));
	687	+ if (ret)
	688	+ goto out;
	689	+ downgrade_write(&device->client_data_rwsem);
	690	+ if (client->add) {
	691	+ if (client->add(device)) {
	692	+ /*
	693	+ * If a client fails to add then the error code is
	694	+ * ignored, but we won't call any more ops on this
	695	+ * client.
	696	+ */
	697	+ xa_erase(&device->client_data, client->client_id);
	698	+ up_read(&device->client_data_rwsem);
	699	+ ib_device_put(device);
	700	+ ib_client_put(client);
	701	+ return 0;
	702	+ }
	703	+ }
	704	+
	705	+ /* Readers shall not see a client until add has been completed */
	706	+ xa_set_mark(&device->client_data, client->client_id,
	707	+ CLIENT_DATA_REGISTERED);
	708	+ up_read(&device->client_data_rwsem);
	709	+ return 0;
	710	+
	711	+out:
	712	+ ib_device_put(device);
	713	+ ib_client_put(client);
	714	+out_unlock:
	715	+ up_write(&device->client_data_rwsem);
	716	+ return ret;
	717	+}
	718	+
	719	+static void remove_client_context(struct ib_device *device,
	720	+ unsigned int client_id)
	721	+{
	722	+ struct ib_client *client;
	723	+ void *client_data;
	724	+
	725	+ down_write(&device->client_data_rwsem);
	726	+ if (!xa_get_mark(&device->client_data, client_id,
	727	+ CLIENT_DATA_REGISTERED)) {
	728	+ up_write(&device->client_data_rwsem);
	729	+ return;
	730	+ }
	731	+ client_data = xa_load(&device->client_data, client_id);
	732	+ xa_clear_mark(&device->client_data, client_id, CLIENT_DATA_REGISTERED);
	733	+ client = xa_load(&clients, client_id);
	734	+ up_write(&device->client_data_rwsem);
	735	+
	736	+ /*
	737	+ * Notice we cannot be holding any exclusive locks when calling the
	738	+ * remove callback as the remove callback can recurse back into any
	739	+ * public functions in this module and thus try for any locks those
	740	+ * functions take.
	741	+ *
	742	+ * For this reason clients and drivers should not call the
	743	+ * unregistration functions will holdling any locks.
	744	+ */
	745	+ if (client->remove)
	746	+ client->remove(device, client_data);
	747	+
	748	+ xa_erase(&device->client_data, client_id);
	749	+ ib_device_put(device);
	750	+ ib_client_put(client);
	751	+}
	752	+
	753	+static int alloc_port_data(struct ib_device *device)
	754	+{
	755	+ struct ib_port_data_rcu *pdata_rcu;
	756	+ unsigned int port;
	757	+
	758	+ if (device->port_data)
	759	+ return 0;
	760	+
	761	+ /* This can only be called once the physical port range is defined */
	762	+ if (WARN_ON(!device->phys_port_cnt))
	763	+ return -EINVAL;
	764	+
	765	+ /*
	766	+ * device->port_data is indexed directly by the port number to make
	767	+ * access to this data as efficient as possible.
	768	+ *
	769	+ * Therefore port_data is declared as a 1 based array with potential
	770	+ * empty slots at the beginning.
	771	+ */
	772	+ pdata_rcu = kzalloc(struct_size(pdata_rcu, pdata,
	773	+ rdma_end_port(device) + 1),
	774	+ GFP_KERNEL);
	775	+ if (!pdata_rcu)
303	776	return -ENOMEM;
	777	+ /*
	778	+ * The rcu_head is put in front of the port data array and the stored
	779	+ * pointer is adjusted since we never need to see that member until
	780	+ * kfree_rcu.
	781	+ */
	782	+ device->port_data = pdata_rcu->pdata;
304	783
305		- context->client = client;
306		- context->data = NULL;
307		- context->going_down = false;
	784	+ rdma_for_each_port (device, port) {
	785	+ struct ib_port_data *pdata = &device->port_data[port];
308	786
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		-
	787	+ pdata->ib_dev = device;
	788	+ spin_lock_init(&pdata->pkey_list_lock);
	789	+ INIT_LIST_HEAD(&pdata->pkey_list);
	790	+ spin_lock_init(&pdata->netdev_lock);
	791	+ INIT_HLIST_NODE(&pdata->ndev_hash_link);
	792	+ }
315	793	return 0;
316	794	}
317	795
..	..	@@ -321,29 +799,20 @@
321	799	rdma_max_mad_size(dev, port) != 0);
322	800	}
323	801
324		-static int read_port_immutable(struct ib_device *device)
	802	+static int setup_port_data(struct ib_device *device)
325	803	{
	804	+ unsigned int port;
326	805	int ret;
327		- u8 start_port = rdma_start_port(device);
328		- u8 end_port = rdma_end_port(device);
329		- u8 port;
330	806
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;
	807	+ ret = alloc_port_data(device);
	808	+ if (ret)
	809	+ return ret;
343	810
344		- for (port = start_port; port <= end_port; ++port) {
345		- ret = device->get_port_immutable(device, port,
346		- &device->port_immutable[port]);
	811	+ rdma_for_each_port (device, port) {
	812	+ struct ib_port_data *pdata = &device->port_data[port];
	813	+
	814	+ ret = device->ops.get_port_immutable(device, port,
	815	+ &pdata->immutable);
347	816	if (ret)
348	817	return ret;
349	818
..	..	@@ -355,46 +824,23 @@
355	824
356	825	void ib_get_device_fw_str(struct ib_device dev, char str)
357	826	{
358		- if (dev->get_dev_fw_str)
359		- dev->get_dev_fw_str(dev, str);
	827	+ if (dev->ops.get_dev_fw_str)
	828	+ dev->ops.get_dev_fw_str(dev, str);
360	829	else
361	830	str[0] = '\0';
362	831	}
363	832	EXPORT_SYMBOL(ib_get_device_fw_str);
364	833
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	834	static void ib_policy_change_task(struct work_struct *work)
390	835	{
391	836	struct ib_device *dev;
	837	+ unsigned long index;
392	838
393		- down_read(&lists_rwsem);
394		- list_for_each_entry(dev, &device_list, core_list) {
395		- int i;
	839	+ down_read(&devices_rwsem);
	840	+ xa_for_each_marked (&devices, index, dev, DEVICE_REGISTERED) {
	841	+ unsigned int i;
396	842
397		- for (i = rdma_start_port(dev); i <= rdma_end_port(dev); i++) {
	843	+ rdma_for_each_port (dev, i) {
398	844	u64 sp;
399	845	int ret = ib_get_cached_subnet_prefix(dev,
400	846	i,
..	..	@@ -407,7 +853,7 @@
407	853	ib_security_cache_change(dev, i, sp);
408	854	}
409	855	}
410		- up_read(&lists_rwsem);
	856	+ up_read(&devices_rwsem);
411	857	}
412	858
413	859	static int ib_security_change(struct notifier_block *nb, unsigned long event,
..	..	@@ -417,208 +863,857 @@
417	863	return NOTIFY_DONE;
418	864
419	865	schedule_work(&ib_policy_change_work);
	866	+ ib_mad_agent_security_change();
420	867
421	868	return NOTIFY_OK;
422	869	}
423	870
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)
	871	+static void compatdev_release(struct device *dev)
432	872	{
	873	+ struct ib_core_device *cdev =
	874	+ container_of(dev, struct ib_core_device, dev);
	875	+
	876	+ kfree(cdev);
	877	+}
	878	+
	879	+static int add_one_compat_dev(struct ib_device *device,
	880	+ struct rdma_dev_net *rnet)
	881	+{
	882	+ struct ib_core_device *cdev;
	883	+ int ret;
	884	+
	885	+ lockdep_assert_held(&rdma_nets_rwsem);
	886	+ if (!ib_devices_shared_netns)
	887	+ return 0;
	888	+
433	889	/*
434		- * The device index to allow stable naming.
435		- * Similar to struct net -> ifindex.
	890	+ * Create and add compat device in all namespaces other than where it
	891	+ * is currently bound to.
436	892	*/
437		- static u32 index;
	893	+ if (net_eq(read_pnet(&rnet->net),
	894	+ read_pnet(&device->coredev.rdma_net)))
	895	+ return 0;
438	896
439		- for (;;) {
440		- if (!(++index))
441		- index = 1;
442		-
443		- if (!__ib_device_get_by_index(index))
444		- return index;
	897	+ /*
	898	+ * The first of init_net() or ib_register_device() to take the
	899	+ * compat_devs_mutex wins and gets to add the device. Others will wait
	900	+ * for completion here.
	901	+ */
	902	+ mutex_lock(&device->compat_devs_mutex);
	903	+ cdev = xa_load(&device->compat_devs, rnet->id);
	904	+ if (cdev) {
	905	+ ret = 0;
	906	+ goto done;
445	907	}
	908	+ ret = xa_reserve(&device->compat_devs, rnet->id, GFP_KERNEL);
	909	+ if (ret)
	910	+ goto done;
	911	+
	912	+ cdev = kzalloc(sizeof(*cdev), GFP_KERNEL);
	913	+ if (!cdev) {
	914	+ ret = -ENOMEM;
	915	+ goto cdev_err;
	916	+ }
	917	+
	918	+ cdev->dev.parent = device->dev.parent;
	919	+ rdma_init_coredev(cdev, device, read_pnet(&rnet->net));
	920	+ cdev->dev.release = compatdev_release;
	921	+ ret = dev_set_name(&cdev->dev, "%s", dev_name(&device->dev));
	922	+ if (ret)
	923	+ goto add_err;
	924	+
	925	+ ret = device_add(&cdev->dev);
	926	+ if (ret)
	927	+ goto add_err;
	928	+ ret = ib_setup_port_attrs(cdev);
	929	+ if (ret)
	930	+ goto port_err;
	931	+
	932	+ ret = xa_err(xa_store(&device->compat_devs, rnet->id,
	933	+ cdev, GFP_KERNEL));
	934	+ if (ret)
	935	+ goto insert_err;
	936	+
	937	+ mutex_unlock(&device->compat_devs_mutex);
	938	+ return 0;
	939	+
	940	+insert_err:
	941	+ ib_free_port_attrs(cdev);
	942	+port_err:
	943	+ device_del(&cdev->dev);
	944	+add_err:
	945	+ put_device(&cdev->dev);
	946	+cdev_err:
	947	+ xa_release(&device->compat_devs, rnet->id);
	948	+done:
	949	+ mutex_unlock(&device->compat_devs_mutex);
	950	+ return ret;
	951	+}
	952	+
	953	+static void remove_one_compat_dev(struct ib_device *device, u32 id)
	954	+{
	955	+ struct ib_core_device *cdev;
	956	+
	957	+ mutex_lock(&device->compat_devs_mutex);
	958	+ cdev = xa_erase(&device->compat_devs, id);
	959	+ mutex_unlock(&device->compat_devs_mutex);
	960	+ if (cdev) {
	961	+ ib_free_port_attrs(cdev);
	962	+ device_del(&cdev->dev);
	963	+ put_device(&cdev->dev);
	964	+ }
	965	+}
	966	+
	967	+static void remove_compat_devs(struct ib_device *device)
	968	+{
	969	+ struct ib_core_device *cdev;
	970	+ unsigned long index;
	971	+
	972	+ xa_for_each (&device->compat_devs, index, cdev)
	973	+ remove_one_compat_dev(device, index);
	974	+}
	975	+
	976	+static int add_compat_devs(struct ib_device *device)
	977	+{
	978	+ struct rdma_dev_net *rnet;
	979	+ unsigned long index;
	980	+ int ret = 0;
	981	+
	982	+ lockdep_assert_held(&devices_rwsem);
	983	+
	984	+ down_read(&rdma_nets_rwsem);
	985	+ xa_for_each (&rdma_nets, index, rnet) {
	986	+ ret = add_one_compat_dev(device, rnet);
	987	+ if (ret)
	988	+ break;
	989	+ }
	990	+ up_read(&rdma_nets_rwsem);
	991	+ return ret;
	992	+}
	993	+
	994	+static void remove_all_compat_devs(void)
	995	+{
	996	+ struct ib_compat_device *cdev;
	997	+ struct ib_device *dev;
	998	+ unsigned long index;
	999	+
	1000	+ down_read(&devices_rwsem);
	1001	+ xa_for_each (&devices, index, dev) {
	1002	+ unsigned long c_index = 0;
	1003	+
	1004	+ /* Hold nets_rwsem so that any other thread modifying this
	1005	+ * system param can sync with this thread.
	1006	+ */
	1007	+ down_read(&rdma_nets_rwsem);
	1008	+ xa_for_each (&dev->compat_devs, c_index, cdev)
	1009	+ remove_one_compat_dev(dev, c_index);
	1010	+ up_read(&rdma_nets_rwsem);
	1011	+ }
	1012	+ up_read(&devices_rwsem);
	1013	+}
	1014	+
	1015	+static int add_all_compat_devs(void)
	1016	+{
	1017	+ struct rdma_dev_net *rnet;
	1018	+ struct ib_device *dev;
	1019	+ unsigned long index;
	1020	+ int ret = 0;
	1021	+
	1022	+ down_read(&devices_rwsem);
	1023	+ xa_for_each_marked (&devices, index, dev, DEVICE_REGISTERED) {
	1024	+ unsigned long net_index = 0;
	1025	+
	1026	+ /* Hold nets_rwsem so that any other thread modifying this
	1027	+ * system param can sync with this thread.
	1028	+ */
	1029	+ down_read(&rdma_nets_rwsem);
	1030	+ xa_for_each (&rdma_nets, net_index, rnet) {
	1031	+ ret = add_one_compat_dev(dev, rnet);
	1032	+ if (ret)
	1033	+ break;
	1034	+ }
	1035	+ up_read(&rdma_nets_rwsem);
	1036	+ }
	1037	+ up_read(&devices_rwsem);
	1038	+ if (ret)
	1039	+ remove_all_compat_devs();
	1040	+ return ret;
	1041	+}
	1042	+
	1043	+int rdma_compatdev_set(u8 enable)
	1044	+{
	1045	+ struct rdma_dev_net *rnet;
	1046	+ unsigned long index;
	1047	+ int ret = 0;
	1048	+
	1049	+ down_write(&rdma_nets_rwsem);
	1050	+ if (ib_devices_shared_netns == enable) {
	1051	+ up_write(&rdma_nets_rwsem);
	1052	+ return 0;
	1053	+ }
	1054	+
	1055	+ /* enable/disable of compat devices is not supported
	1056	+ * when more than default init_net exists.
	1057	+ */
	1058	+ xa_for_each (&rdma_nets, index, rnet) {
	1059	+ ret++;
	1060	+ break;
	1061	+ }
	1062	+ if (!ret)
	1063	+ ib_devices_shared_netns = enable;
	1064	+ up_write(&rdma_nets_rwsem);
	1065	+ if (ret)
	1066	+ return -EBUSY;
	1067	+
	1068	+ if (enable)
	1069	+ ret = add_all_compat_devs();
	1070	+ else
	1071	+ remove_all_compat_devs();
	1072	+ return ret;
	1073	+}
	1074	+
	1075	+static void rdma_dev_exit_net(struct net *net)
	1076	+{
	1077	+ struct rdma_dev_net *rnet = rdma_net_to_dev_net(net);
	1078	+ struct ib_device *dev;
	1079	+ unsigned long index;
	1080	+ int ret;
	1081	+
	1082	+ down_write(&rdma_nets_rwsem);
	1083	+ /*
	1084	+ * Prevent the ID from being re-used and hide the id from xa_for_each.
	1085	+ */
	1086	+ ret = xa_err(xa_store(&rdma_nets, rnet->id, NULL, GFP_KERNEL));
	1087	+ WARN_ON(ret);
	1088	+ up_write(&rdma_nets_rwsem);
	1089	+
	1090	+ down_read(&devices_rwsem);
	1091	+ xa_for_each (&devices, index, dev) {
	1092	+ get_device(&dev->dev);
	1093	+ /*
	1094	+ * Release the devices_rwsem so that pontentially blocking
	1095	+ * device_del, doesn't hold the devices_rwsem for too long.
	1096	+ */
	1097	+ up_read(&devices_rwsem);
	1098	+
	1099	+ remove_one_compat_dev(dev, rnet->id);
	1100	+
	1101	+ /*
	1102	+ * If the real device is in the NS then move it back to init.
	1103	+ */
	1104	+ rdma_dev_change_netns(dev, net, &init_net);
	1105	+
	1106	+ put_device(&dev->dev);
	1107	+ down_read(&devices_rwsem);
	1108	+ }
	1109	+ up_read(&devices_rwsem);
	1110	+
	1111	+ rdma_nl_net_exit(rnet);
	1112	+ xa_erase(&rdma_nets, rnet->id);
	1113	+}
	1114	+
	1115	+static __net_init int rdma_dev_init_net(struct net *net)
	1116	+{
	1117	+ struct rdma_dev_net *rnet = rdma_net_to_dev_net(net);
	1118	+ unsigned long index;
	1119	+ struct ib_device *dev;
	1120	+ int ret;
	1121	+
	1122	+ write_pnet(&rnet->net, net);
	1123	+
	1124	+ ret = rdma_nl_net_init(rnet);
	1125	+ if (ret)
	1126	+ return ret;
	1127	+
	1128	+ /* No need to create any compat devices in default init_net. */
	1129	+ if (net_eq(net, &init_net))
	1130	+ return 0;
	1131	+
	1132	+ ret = xa_alloc(&rdma_nets, &rnet->id, rnet, xa_limit_32b, GFP_KERNEL);
	1133	+ if (ret) {
	1134	+ rdma_nl_net_exit(rnet);
	1135	+ return ret;
	1136	+ }
	1137	+
	1138	+ down_read(&devices_rwsem);
	1139	+ xa_for_each_marked (&devices, index, dev, DEVICE_REGISTERED) {
	1140	+ /* Hold nets_rwsem so that netlink command cannot change
	1141	+ * system configuration for device sharing mode.
	1142	+ */
	1143	+ down_read(&rdma_nets_rwsem);
	1144	+ ret = add_one_compat_dev(dev, rnet);
	1145	+ up_read(&rdma_nets_rwsem);
	1146	+ if (ret)
	1147	+ break;
	1148	+ }
	1149	+ up_read(&devices_rwsem);
	1150	+
	1151	+ if (ret)
	1152	+ rdma_dev_exit_net(net);
	1153	+
	1154	+ return ret;
	1155	+}
	1156	+
	1157	+/*
	1158	+ * Assign the unique string device name and the unique device index. This is
	1159	+ * undone by ib_dealloc_device.
	1160	+ */
	1161	+static int assign_name(struct ib_device device, const char name)
	1162	+{
	1163	+ static u32 last_id;
	1164	+ int ret;
	1165	+
	1166	+ down_write(&devices_rwsem);
	1167	+ /* Assign a unique name to the device */
	1168	+ if (strchr(name, '%'))
	1169	+ ret = alloc_name(device, name);
	1170	+ else
	1171	+ ret = dev_set_name(&device->dev, name);
	1172	+ if (ret)
	1173	+ goto out;
	1174	+
	1175	+ if (__ib_device_get_by_name(dev_name(&device->dev))) {
	1176	+ ret = -ENFILE;
	1177	+ goto out;
	1178	+ }
	1179	+ strlcpy(device->name, dev_name(&device->dev), IB_DEVICE_NAME_MAX);
	1180	+
	1181	+ ret = xa_alloc_cyclic(&devices, &device->index, device, xa_limit_31b,
	1182	+ &last_id, GFP_KERNEL);
	1183	+ if (ret > 0)
	1184	+ ret = 0;
	1185	+
	1186	+out:
	1187	+ up_write(&devices_rwsem);
	1188	+ return ret;
	1189	+}
	1190	+
	1191	+/*
	1192	+ * setup_device() allocates memory and sets up data that requires calling the
	1193	+ * device ops, this is the only reason these actions are not done during
	1194	+ * ib_alloc_device. It is undone by ib_dealloc_device().
	1195	+ */
	1196	+static int setup_device(struct ib_device *device)
	1197	+{
	1198	+ struct ib_udata uhw = {.outlen = 0, .inlen = 0};
	1199	+ int ret;
	1200	+
	1201	+ ib_device_check_mandatory(device);
	1202	+
	1203	+ ret = setup_port_data(device);
	1204	+ if (ret) {
	1205	+ dev_warn(&device->dev, "Couldn't create per-port data\n");
	1206	+ return ret;
	1207	+ }
	1208	+
	1209	+ memset(&device->attrs, 0, sizeof(device->attrs));
	1210	+ ret = device->ops.query_device(device, &device->attrs, &uhw);
	1211	+ if (ret) {
	1212	+ dev_warn(&device->dev,
	1213	+ "Couldn't query the device attributes\n");
	1214	+ return ret;
	1215	+ }
	1216	+
	1217	+ return 0;
	1218	+}
	1219	+
	1220	+static void disable_device(struct ib_device *device)
	1221	+{
	1222	+ u32 cid;
	1223	+
	1224	+ WARN_ON(!refcount_read(&device->refcount));
	1225	+
	1226	+ down_write(&devices_rwsem);
	1227	+ xa_clear_mark(&devices, device->index, DEVICE_REGISTERED);
	1228	+ up_write(&devices_rwsem);
	1229	+
	1230	+ /*
	1231	+ * Remove clients in LIFO order, see assign_client_id. This could be
	1232	+ * more efficient if xarray learns to reverse iterate. Since no new
	1233	+ * clients can be added to this ib_device past this point we only need
	1234	+ * the maximum possible client_id value here.
	1235	+ */
	1236	+ down_read(&clients_rwsem);
	1237	+ cid = highest_client_id;
	1238	+ up_read(&clients_rwsem);
	1239	+ while (cid) {
	1240	+ cid--;
	1241	+ remove_client_context(device, cid);
	1242	+ }
	1243	+
	1244	+ ib_cq_pool_destroy(device);
	1245	+
	1246	+ /* Pairs with refcount_set in enable_device */
	1247	+ ib_device_put(device);
	1248	+ wait_for_completion(&device->unreg_completion);
	1249	+
	1250	+ /*
	1251	+ * compat devices must be removed after device refcount drops to zero.
	1252	+ * Otherwise init_net() may add more compatdevs after removing compat
	1253	+ * devices and before device is disabled.
	1254	+ */
	1255	+ remove_compat_devs(device);
	1256	+}
	1257	+
	1258	+/*
	1259	+ * An enabled device is visible to all clients and to all the public facing
	1260	+ * APIs that return a device pointer. This always returns with a new get, even
	1261	+ * if it fails.
	1262	+ */
	1263	+static int enable_device_and_get(struct ib_device *device)
	1264	+{
	1265	+ struct ib_client *client;
	1266	+ unsigned long index;
	1267	+ int ret = 0;
	1268	+
	1269	+ /*
	1270	+ * One ref belongs to the xa and the other belongs to this
	1271	+ * thread. This is needed to guard against parallel unregistration.
	1272	+ */
	1273	+ refcount_set(&device->refcount, 2);
	1274	+ down_write(&devices_rwsem);
	1275	+ xa_set_mark(&devices, device->index, DEVICE_REGISTERED);
	1276	+
	1277	+ /*
	1278	+ * By using downgrade_write() we ensure that no other thread can clear
	1279	+ * DEVICE_REGISTERED while we are completing the client setup.
	1280	+ */
	1281	+ downgrade_write(&devices_rwsem);
	1282	+
	1283	+ if (device->ops.enable_driver) {
	1284	+ ret = device->ops.enable_driver(device);
	1285	+ if (ret)
	1286	+ goto out;
	1287	+ }
	1288	+
	1289	+ ib_cq_pool_init(device);
	1290	+
	1291	+ down_read(&clients_rwsem);
	1292	+ xa_for_each_marked (&clients, index, client, CLIENT_REGISTERED) {
	1293	+ ret = add_client_context(device, client);
	1294	+ if (ret)
	1295	+ break;
	1296	+ }
	1297	+ up_read(&clients_rwsem);
	1298	+ if (!ret)
	1299	+ ret = add_compat_devs(device);
	1300	+out:
	1301	+ up_read(&devices_rwsem);
	1302	+ return ret;
	1303	+}
	1304	+
	1305	+static void prevent_dealloc_device(struct ib_device *ib_dev)
	1306	+{
446	1307	}
447	1308
448	1309	/**
449	1310	* ib_register_device - Register an IB device with IB core
450		- * @device:Device to register
	1311	+ * @device: Device to register
	1312	+ * @name: unique string device name. This may include a '%' which will
	1313	+ * cause a unique index to be added to the passed device name.
	1314	+ * @dma_device: pointer to a DMA-capable device. If %NULL, then the IB
	1315	+ * device will be used. In this case the caller should fully
	1316	+ * setup the ibdev for DMA. This usually means using dma_virt_ops.
451	1317	*
452	1318	* Low-level drivers use ib_register_device() to register their
453	1319	* devices with the IB core. All registered clients will receive a
454	1320	* callback for each device that is added. @device must be allocated
455	1321	* with ib_alloc_device().
	1322	+ *
	1323	+ * If the driver uses ops.dealloc_driver and calls any ib_unregister_device()
	1324	+ * asynchronously then the device pointer may become freed as soon as this
	1325	+ * function returns.
456	1326	*/
457		-int ib_register_device(struct ib_device *device,
458		- int (port_callback)(struct ib_device ,
459		- u8, struct kobject *))
	1327	+int ib_register_device(struct ib_device device, const char name,
	1328	+ struct device *dma_device)
460	1329	{
461	1330	int ret;
462		- struct ib_client *client;
463		- struct ib_udata uhw = {.outlen = 0, .inlen = 0};
464		- struct device *parent = device->dev.parent;
465	1331
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		- }
	1332	+ ret = assign_name(device, name);
	1333	+ if (ret)
	1334	+ return ret;
495	1335
496		- mutex_lock(&device_mutex);
	1336	+ /*
	1337	+ * If the caller does not provide a DMA capable device then the IB core
	1338	+ * will set up ib_sge and scatterlist structures that stash the kernel
	1339	+ * virtual address into the address field.
	1340	+ */
	1341	+ WARN_ON(dma_device && !dma_device->dma_parms);
	1342	+ device->dma_device = dma_device;
497	1343
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		- }
	1344	+ ret = setup_device(device);
	1345	+ if (ret)
	1346	+ return ret;
521	1347
522	1348	ret = ib_cache_setup_one(device);
523	1349	if (ret) {
524		- pr_warn("Couldn't set up InfiniBand P_Key/GID cache\n");
525		- goto port_cleanup;
	1350	+ dev_warn(&device->dev,
	1351	+ "Couldn't set up InfiniBand P_Key/GID cache\n");
	1352	+ return ret;
526	1353	}
527	1354
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		- }
	1355	+ ib_device_register_rdmacg(device);
533	1356
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");
	1357	+ rdma_counter_init(device);
	1358	+
	1359	+ /*
	1360	+ * Ensure that ADD uevent is not fired because it
	1361	+ * is too early amd device is not initialized yet.
	1362	+ */
	1363	+ dev_set_uevent_suppress(&device->dev, true);
	1364	+ ret = device_add(&device->dev);
	1365	+ if (ret)
538	1366	goto cg_cleanup;
539		- }
540	1367
541		- ret = ib_device_register_sysfs(device, port_callback);
	1368	+ ret = ib_device_register_sysfs(device);
542	1369	if (ret) {
543		- pr_warn("Couldn't register device %s with driver model\n",
544		- device->name);
545		- goto cg_cleanup;
	1370	+ dev_warn(&device->dev,
	1371	+ "Couldn't register device with driver model\n");
	1372	+ goto dev_cleanup;
546	1373	}
547	1374
548		- device->reg_state = IB_DEV_REGISTERED;
	1375	+ ret = enable_device_and_get(device);
	1376	+ if (ret) {
	1377	+ void (dealloc_fn)(struct ib_device );
549	1378
550		- list_for_each_entry(client, &client_list, list)
551		- if (!add_client_context(device, client) && client->add)
552		- client->add(device);
	1379	+ /*
	1380	+ * If we hit this error flow then we don't want to
	1381	+ * automatically dealloc the device since the caller is
	1382	+ * expected to call ib_dealloc_device() after
	1383	+ * ib_register_device() fails. This is tricky due to the
	1384	+ * possibility for a parallel unregistration along with this
	1385	+ * error flow. Since we have a refcount here we know any
	1386	+ * parallel flow is stopped in disable_device and will see the
	1387	+ * special dealloc_driver pointer, causing the responsibility to
	1388	+ * ib_dealloc_device() to revert back to this thread.
	1389	+ */
	1390	+ dealloc_fn = device->ops.dealloc_driver;
	1391	+ device->ops.dealloc_driver = prevent_dealloc_device;
	1392	+ ib_device_put(device);
	1393	+ __ib_unregister_device(device);
	1394	+ device->ops.dealloc_driver = dealloc_fn;
	1395	+ dev_set_uevent_suppress(&device->dev, false);
	1396	+ return ret;
	1397	+ }
	1398	+ dev_set_uevent_suppress(&device->dev, false);
	1399	+ /* Mark for userspace that device is ready */
	1400	+ kobject_uevent(&device->dev.kobj, KOBJ_ADD);
	1401	+ ib_device_put(device);
553	1402
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	1403	return 0;
560	1404
	1405	+dev_cleanup:
	1406	+ device_del(&device->dev);
561	1407	cg_cleanup:
	1408	+ dev_set_uevent_suppress(&device->dev, false);
562	1409	ib_device_unregister_rdmacg(device);
563		-cache_cleanup:
564	1410	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	1411	return ret;
571	1412	}
572	1413	EXPORT_SYMBOL(ib_register_device);
573	1414
	1415	+/* Callers must hold a get on the device. */
	1416	+static void __ib_unregister_device(struct ib_device *ib_dev)
	1417	+{
	1418	+ /*
	1419	+ * We have a registration lock so that all the calls to unregister are
	1420	+ * fully fenced, once any unregister returns the device is truely
	1421	+ * unregistered even if multiple callers are unregistering it at the
	1422	+ * same time. This also interacts with the registration flow and
	1423	+ * provides sane semantics if register and unregister are racing.
	1424	+ */
	1425	+ mutex_lock(&ib_dev->unregistration_lock);
	1426	+ if (!refcount_read(&ib_dev->refcount))
	1427	+ goto out;
	1428	+
	1429	+ disable_device(ib_dev);
	1430	+
	1431	+ /* Expedite removing unregistered pointers from the hash table */
	1432	+ free_netdevs(ib_dev);
	1433	+
	1434	+ ib_device_unregister_sysfs(ib_dev);
	1435	+ device_del(&ib_dev->dev);
	1436	+ ib_device_unregister_rdmacg(ib_dev);
	1437	+ ib_cache_cleanup_one(ib_dev);
	1438	+
	1439	+ /*
	1440	+ * Drivers using the new flow may not call ib_dealloc_device except
	1441	+ * in error unwind prior to registration success.
	1442	+ */
	1443	+ if (ib_dev->ops.dealloc_driver &&
	1444	+ ib_dev->ops.dealloc_driver != prevent_dealloc_device) {
	1445	+ WARN_ON(kref_read(&ib_dev->dev.kobj.kref) <= 1);
	1446	+ ib_dealloc_device(ib_dev);
	1447	+ }
	1448	+out:
	1449	+ mutex_unlock(&ib_dev->unregistration_lock);
	1450	+}
	1451	+
574	1452	/**
575	1453	* ib_unregister_device - Unregister an IB device
576		- * @device:Device to unregister
	1454	+ * @ib_dev: The device to unregister
577	1455	*
578	1456	* Unregister an IB device. All clients will receive a remove callback.
	1457	+ *
	1458	+ * Callers should call this routine only once, and protect against races with
	1459	+ * registration. Typically it should only be called as part of a remove
	1460	+ * callback in an implementation of driver core's struct device_driver and
	1461	+ * related.
	1462	+ *
	1463	+ * If ops.dealloc_driver is used then ib_dev will be freed upon return from
	1464	+ * this function.
579	1465	*/
580		-void ib_unregister_device(struct ib_device *device)
	1466	+void ib_unregister_device(struct ib_device *ib_dev)
581	1467	{
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;
	1468	+ get_device(&ib_dev->dev);
	1469	+ __ib_unregister_device(ib_dev);
	1470	+ put_device(&ib_dev->dev);
620	1471	}
621	1472	EXPORT_SYMBOL(ib_unregister_device);
	1473	+
	1474	+/**
	1475	+ * ib_unregister_device_and_put - Unregister a device while holding a 'get'
	1476	+ * @ib_dev: The device to unregister
	1477	+ *
	1478	+ * This is the same as ib_unregister_device(), except it includes an internal
	1479	+ * ib_device_put() that should match a 'get' obtained by the caller.
	1480	+ *
	1481	+ * It is safe to call this routine concurrently from multiple threads while
	1482	+ * holding the 'get'. When the function returns the device is fully
	1483	+ * unregistered.
	1484	+ *
	1485	+ * Drivers using this flow MUST use the driver_unregister callback to clean up
	1486	+ * their resources associated with the device and dealloc it.
	1487	+ */
	1488	+void ib_unregister_device_and_put(struct ib_device *ib_dev)
	1489	+{
	1490	+ WARN_ON(!ib_dev->ops.dealloc_driver);
	1491	+ get_device(&ib_dev->dev);
	1492	+ ib_device_put(ib_dev);
	1493	+ __ib_unregister_device(ib_dev);
	1494	+ put_device(&ib_dev->dev);
	1495	+}
	1496	+EXPORT_SYMBOL(ib_unregister_device_and_put);
	1497	+
	1498	+/**
	1499	+ * ib_unregister_driver - Unregister all IB devices for a driver
	1500	+ * @driver_id: The driver to unregister
	1501	+ *
	1502	+ * This implements a fence for device unregistration. It only returns once all
	1503	+ * devices associated with the driver_id have fully completed their
	1504	+ * unregistration and returned from ib_unregister_device*().
	1505	+ *
	1506	+ * If device's are not yet unregistered it goes ahead and starts unregistering
	1507	+ * them.
	1508	+ *
	1509	+ * This does not block creation of new devices with the given driver_id, that
	1510	+ * is the responsibility of the caller.
	1511	+ */
	1512	+void ib_unregister_driver(enum rdma_driver_id driver_id)
	1513	+{
	1514	+ struct ib_device *ib_dev;
	1515	+ unsigned long index;
	1516	+
	1517	+ down_read(&devices_rwsem);
	1518	+ xa_for_each (&devices, index, ib_dev) {
	1519	+ if (ib_dev->ops.driver_id != driver_id)
	1520	+ continue;
	1521	+
	1522	+ get_device(&ib_dev->dev);
	1523	+ up_read(&devices_rwsem);
	1524	+
	1525	+ WARN_ON(!ib_dev->ops.dealloc_driver);
	1526	+ __ib_unregister_device(ib_dev);
	1527	+
	1528	+ put_device(&ib_dev->dev);
	1529	+ down_read(&devices_rwsem);
	1530	+ }
	1531	+ up_read(&devices_rwsem);
	1532	+}
	1533	+EXPORT_SYMBOL(ib_unregister_driver);
	1534	+
	1535	+static void ib_unregister_work(struct work_struct *work)
	1536	+{
	1537	+ struct ib_device *ib_dev =
	1538	+ container_of(work, struct ib_device, unregistration_work);
	1539	+
	1540	+ __ib_unregister_device(ib_dev);
	1541	+ put_device(&ib_dev->dev);
	1542	+}
	1543	+
	1544	+/**
	1545	+ * ib_unregister_device_queued - Unregister a device using a work queue
	1546	+ * @ib_dev: The device to unregister
	1547	+ *
	1548	+ * This schedules an asynchronous unregistration using a WQ for the device. A
	1549	+ * driver should use this to avoid holding locks while doing unregistration,
	1550	+ * such as holding the RTNL lock.
	1551	+ *
	1552	+ * Drivers using this API must use ib_unregister_driver before module unload
	1553	+ * to ensure that all scheduled unregistrations have completed.
	1554	+ */
	1555	+void ib_unregister_device_queued(struct ib_device *ib_dev)
	1556	+{
	1557	+ WARN_ON(!refcount_read(&ib_dev->refcount));
	1558	+ WARN_ON(!ib_dev->ops.dealloc_driver);
	1559	+ get_device(&ib_dev->dev);
	1560	+ if (!queue_work(system_unbound_wq, &ib_dev->unregistration_work))
	1561	+ put_device(&ib_dev->dev);
	1562	+}
	1563	+EXPORT_SYMBOL(ib_unregister_device_queued);
	1564	+
	1565	+/*
	1566	+ * The caller must pass in a device that has the kref held and the refcount
	1567	+ * released. If the device is in cur_net and still registered then it is moved
	1568	+ * into net.
	1569	+ */
	1570	+static int rdma_dev_change_netns(struct ib_device device, struct net cur_net,
	1571	+ struct net *net)
	1572	+{
	1573	+ int ret2 = -EINVAL;
	1574	+ int ret;
	1575	+
	1576	+ mutex_lock(&device->unregistration_lock);
	1577	+
	1578	+ /*
	1579	+ * If a device not under ib_device_get() or if the unregistration_lock
	1580	+ * is not held, the namespace can be changed, or it can be unregistered.
	1581	+ * Check again under the lock.
	1582	+ */
	1583	+ if (refcount_read(&device->refcount) == 0 \|\|
	1584	+ !net_eq(cur_net, read_pnet(&device->coredev.rdma_net))) {
	1585	+ ret = -ENODEV;
	1586	+ goto out;
	1587	+ }
	1588	+
	1589	+ kobject_uevent(&device->dev.kobj, KOBJ_REMOVE);
	1590	+ disable_device(device);
	1591	+
	1592	+ /*
	1593	+ * At this point no one can be using the device, so it is safe to
	1594	+ * change the namespace.
	1595	+ */
	1596	+ write_pnet(&device->coredev.rdma_net, net);
	1597	+
	1598	+ down_read(&devices_rwsem);
	1599	+ /*
	1600	+ * Currently rdma devices are system wide unique. So the device name
	1601	+ * is guaranteed free in the new namespace. Publish the new namespace
	1602	+ * at the sysfs level.
	1603	+ */
	1604	+ ret = device_rename(&device->dev, dev_name(&device->dev));
	1605	+ up_read(&devices_rwsem);
	1606	+ if (ret) {
	1607	+ dev_warn(&device->dev,
	1608	+ "%s: Couldn't rename device after namespace change\n",
	1609	+ __func__);
	1610	+ /* Try and put things back and re-enable the device */
	1611	+ write_pnet(&device->coredev.rdma_net, cur_net);
	1612	+ }
	1613	+
	1614	+ ret2 = enable_device_and_get(device);
	1615	+ if (ret2) {
	1616	+ /*
	1617	+ * This shouldn't really happen, but if it does, let the user
	1618	+ * retry at later point. So don't disable the device.
	1619	+ */
	1620	+ dev_warn(&device->dev,
	1621	+ "%s: Couldn't re-enable device after namespace change\n",
	1622	+ __func__);
	1623	+ }
	1624	+ kobject_uevent(&device->dev.kobj, KOBJ_ADD);
	1625	+
	1626	+ ib_device_put(device);
	1627	+out:
	1628	+ mutex_unlock(&device->unregistration_lock);
	1629	+ if (ret)
	1630	+ return ret;
	1631	+ return ret2;
	1632	+}
	1633	+
	1634	+int ib_device_set_netns_put(struct sk_buff *skb,
	1635	+ struct ib_device *dev, u32 ns_fd)
	1636	+{
	1637	+ struct net *net;
	1638	+ int ret;
	1639	+
	1640	+ net = get_net_ns_by_fd(ns_fd);
	1641	+ if (IS_ERR(net)) {
	1642	+ ret = PTR_ERR(net);
	1643	+ goto net_err;
	1644	+ }
	1645	+
	1646	+ if (!netlink_ns_capable(skb, net->user_ns, CAP_NET_ADMIN)) {
	1647	+ ret = -EPERM;
	1648	+ goto ns_err;
	1649	+ }
	1650	+
	1651	+ /*
	1652	+ * Currently supported only for those providers which support
	1653	+ * disassociation and don't do port specific sysfs init. Once a
	1654	+ * port_cleanup infrastructure is implemented, this limitation will be
	1655	+ * removed.
	1656	+ */
	1657	+ if (!dev->ops.disassociate_ucontext \|\| dev->ops.init_port \|\|
	1658	+ ib_devices_shared_netns) {
	1659	+ ret = -EOPNOTSUPP;
	1660	+ goto ns_err;
	1661	+ }
	1662	+
	1663	+ get_device(&dev->dev);
	1664	+ ib_device_put(dev);
	1665	+ ret = rdma_dev_change_netns(dev, current->nsproxy->net_ns, net);
	1666	+ put_device(&dev->dev);
	1667	+
	1668	+ put_net(net);
	1669	+ return ret;
	1670	+
	1671	+ns_err:
	1672	+ put_net(net);
	1673	+net_err:
	1674	+ ib_device_put(dev);
	1675	+ return ret;
	1676	+}
	1677	+
	1678	+static struct pernet_operations rdma_dev_net_ops = {
	1679	+ .init = rdma_dev_init_net,
	1680	+ .exit = rdma_dev_exit_net,
	1681	+ .id = &rdma_dev_net_id,
	1682	+ .size = sizeof(struct rdma_dev_net),
	1683	+};
	1684	+
	1685	+static int assign_client_id(struct ib_client *client)
	1686	+{
	1687	+ int ret;
	1688	+
	1689	+ down_write(&clients_rwsem);
	1690	+ /*
	1691	+ * The add/remove callbacks must be called in FIFO/LIFO order. To
	1692	+ * achieve this we assign client_ids so they are sorted in
	1693	+ * registration order.
	1694	+ */
	1695	+ client->client_id = highest_client_id;
	1696	+ ret = xa_insert(&clients, client->client_id, client, GFP_KERNEL);
	1697	+ if (ret)
	1698	+ goto out;
	1699	+
	1700	+ highest_client_id++;
	1701	+ xa_set_mark(&clients, client->client_id, CLIENT_REGISTERED);
	1702	+
	1703	+out:
	1704	+ up_write(&clients_rwsem);
	1705	+ return ret;
	1706	+}
	1707	+
	1708	+static void remove_client_id(struct ib_client *client)
	1709	+{
	1710	+ down_write(&clients_rwsem);
	1711	+ xa_erase(&clients, client->client_id);
	1712	+ for (; highest_client_id; highest_client_id--)
	1713	+ if (xa_load(&clients, highest_client_id - 1))
	1714	+ break;
	1715	+ up_write(&clients_rwsem);
	1716	+}
622	1717
623	1718	/**
624	1719	* ib_register_client - Register an IB client
..	..	@@ -636,19 +1731,25 @@
636	1731	int ib_register_client(struct ib_client *client)
637	1732	{
638	1733	struct ib_device *device;
	1734	+ unsigned long index;
	1735	+ int ret;
639	1736
640		- mutex_lock(&device_mutex);
	1737	+ refcount_set(&client->uses, 1);
	1738	+ init_completion(&client->uses_zero);
	1739	+ ret = assign_client_id(client);
	1740	+ if (ret)
	1741	+ return ret;
641	1742
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		-
	1743	+ down_read(&devices_rwsem);
	1744	+ xa_for_each_marked (&devices, index, device, DEVICE_REGISTERED) {
	1745	+ ret = add_client_context(device, client);
	1746	+ if (ret) {
	1747	+ up_read(&devices_rwsem);
	1748	+ ib_unregister_client(client);
	1749	+ return ret;
	1750	+ }
	1751	+ }
	1752	+ up_read(&devices_rwsem);
652	1753	return 0;
653	1754	}
654	1755	EXPORT_SYMBOL(ib_register_client);
..	..	@@ -660,80 +1761,140 @@
660	1761	* Upper level users use ib_unregister_client() to remove their client
661	1762	* registration. When ib_unregister_client() is called, the client
662	1763	* will receive a remove callback for each IB device still registered.
	1764	+ *
	1765	+ * This is a full fence, once it returns no client callbacks will be called,
	1766	+ * or are running in another thread.
663	1767	*/
664	1768	void ib_unregister_client(struct ib_client *client)
665	1769	{
666		- struct ib_client_data context, tmp;
667	1770	struct ib_device *device;
668		- unsigned long flags;
	1771	+ unsigned long index;
669	1772
670		- mutex_lock(&device_mutex);
	1773	+ down_write(&clients_rwsem);
	1774	+ ib_client_put(client);
	1775	+ xa_clear_mark(&clients, client->client_id, CLIENT_REGISTERED);
	1776	+ up_write(&clients_rwsem);
671	1777
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);
	1778	+ /* We do not want to have locks while calling client->remove() */
	1779	+ rcu_read_lock();
	1780	+ xa_for_each (&devices, index, device) {
	1781	+ if (!ib_device_try_get(device))
697	1782	continue;
698		- }
	1783	+ rcu_read_unlock();
699	1784
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);
	1785	+ remove_client_context(device, client->client_id);
	1786	+
	1787	+ ib_device_put(device);
	1788	+ rcu_read_lock();
706	1789	}
	1790	+ rcu_read_unlock();
707	1791
708		- mutex_unlock(&device_mutex);
	1792	+ /*
	1793	+ * remove_client_context() is not a fence, it can return even though a
	1794	+ * removal is ongoing. Wait until all removals are completed.
	1795	+ */
	1796	+ wait_for_completion(&client->uses_zero);
	1797	+ remove_client_id(client);
709	1798	}
710	1799	EXPORT_SYMBOL(ib_unregister_client);
711	1800
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)
	1801	+static int __ib_get_global_client_nl_info(const char *client_name,
	1802	+ struct ib_client_nl_info *res)
721	1803	{
722		- struct ib_client_data *context;
723		- void *ret = NULL;
724		- unsigned long flags;
	1804	+ struct ib_client *client;
	1805	+ unsigned long index;
	1806	+ int ret = -ENOENT;
725	1807
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;
	1808	+ down_read(&clients_rwsem);
	1809	+ xa_for_each_marked (&clients, index, client, CLIENT_REGISTERED) {
	1810	+ if (strcmp(client->name, client_name) != 0)
	1811	+ continue;
	1812	+ if (!client->get_global_nl_info) {
	1813	+ ret = -EOPNOTSUPP;
730	1814	break;
731	1815	}
732		- spin_unlock_irqrestore(&device->client_data_lock, flags);
	1816	+ ret = client->get_global_nl_info(res);
	1817	+ if (WARN_ON(ret == -ENOENT))
	1818	+ ret = -EINVAL;
	1819	+ if (!ret && res->cdev)
	1820	+ get_device(res->cdev);
	1821	+ break;
	1822	+ }
	1823	+ up_read(&clients_rwsem);
	1824	+ return ret;
	1825	+}
	1826	+
	1827	+static int __ib_get_client_nl_info(struct ib_device *ibdev,
	1828	+ const char *client_name,
	1829	+ struct ib_client_nl_info *res)
	1830	+{
	1831	+ unsigned long index;
	1832	+ void *client_data;
	1833	+ int ret = -ENOENT;
	1834	+
	1835	+ down_read(&ibdev->client_data_rwsem);
	1836	+ xan_for_each_marked (&ibdev->client_data, index, client_data,
	1837	+ CLIENT_DATA_REGISTERED) {
	1838	+ struct ib_client *client = xa_load(&clients, index);
	1839	+
	1840	+ if (!client \|\| strcmp(client->name, client_name) != 0)
	1841	+ continue;
	1842	+ if (!client->get_nl_info) {
	1843	+ ret = -EOPNOTSUPP;
	1844	+ break;
	1845	+ }
	1846	+ ret = client->get_nl_info(ibdev, client_data, res);
	1847	+ if (WARN_ON(ret == -ENOENT))
	1848	+ ret = -EINVAL;
	1849	+
	1850	+ /*
	1851	+ * The cdev is guaranteed valid as long as we are inside the
	1852	+ * client_data_rwsem as remove_one can't be called. Keep it
	1853	+ * valid for the caller.
	1854	+ */
	1855	+ if (!ret && res->cdev)
	1856	+ get_device(res->cdev);
	1857	+ break;
	1858	+ }
	1859	+ up_read(&ibdev->client_data_rwsem);
733	1860
734	1861	return ret;
735	1862	}
736		-EXPORT_SYMBOL(ib_get_client_data);
	1863	+
	1864	+/**
	1865	+ * ib_get_client_nl_info - Fetch the nl_info from a client
	1866	+ * @device - IB device
	1867	+ * @client_name - Name of the client
	1868	+ * @res - Result of the query
	1869	+ */
	1870	+int ib_get_client_nl_info(struct ib_device ibdev, const char client_name,
	1871	+ struct ib_client_nl_info *res)
	1872	+{
	1873	+ int ret;
	1874	+
	1875	+ if (ibdev)
	1876	+ ret = __ib_get_client_nl_info(ibdev, client_name, res);
	1877	+ else
	1878	+ ret = __ib_get_global_client_nl_info(client_name, res);
	1879	+#ifdef CONFIG_MODULES
	1880	+ if (ret == -ENOENT) {
	1881	+ request_module("rdma-client-%s", client_name);
	1882	+ if (ibdev)
	1883	+ ret = __ib_get_client_nl_info(ibdev, client_name, res);
	1884	+ else
	1885	+ ret = __ib_get_global_client_nl_info(client_name, res);
	1886	+ }
	1887	+#endif
	1888	+ if (ret) {
	1889	+ if (ret == -ENOENT)
	1890	+ return -EOPNOTSUPP;
	1891	+ return ret;
	1892	+ }
	1893	+
	1894	+ if (WARN_ON(!res->cdev))
	1895	+ return -EINVAL;
	1896	+ return 0;
	1897	+}
737	1898
738	1899	/**
739	1900	* ib_set_client_data - Set IB client context
..	..	@@ -741,27 +1902,22 @@
741	1902	* @client:Client to set context for
742	1903	* @data:Context to set
743	1904	*
744		- * ib_set_client_data() sets client context that can be retrieved with
745		- * ib_get_client_data().
	1905	+ * ib_set_client_data() sets client context data that can be retrieved with
	1906	+ * ib_get_client_data(). This can only be called while the client is
	1907	+ * registered to the device, once the ib_client remove() callback returns this
	1908	+ * cannot be called.
746	1909	*/
747	1910	void ib_set_client_data(struct ib_device device, struct ib_client client,
748	1911	void *data)
749	1912	{
750		- struct ib_client_data *context;
751		- unsigned long flags;
	1913	+ void *rc;
752	1914
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		- }
	1915	+ if (WARN_ON(IS_ERR(data)))
	1916	+ data = NULL;
759	1917
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);
	1918	+ rc = xa_store(&device->client_data, client->client_id, data,
	1919	+ GFP_KERNEL);
	1920	+ WARN_ON(xa_is_err(rc));
765	1921	}
766	1922	EXPORT_SYMBOL(ib_set_client_data);
767	1923
..	..	@@ -771,17 +1927,15 @@
771	1927	*
772	1928	* ib_register_event_handler() registers an event handler that will be
773	1929	* 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.
	1930	+ * chapter 11 of the InfiniBand Architecture Specification). This
	1931	+ * callback occurs in workqueue context.
776	1932	*/
777	1933	void ib_register_event_handler(struct ib_event_handler *event_handler)
778	1934	{
779		- unsigned long flags;
780		-
781		- spin_lock_irqsave(&event_handler->device->event_handler_lock, flags);
	1935	+ down_write(&event_handler->device->event_handler_rwsem);
782	1936	list_add_tail(&event_handler->list,
783	1937	&event_handler->device->event_handler_list);
784		- spin_unlock_irqrestore(&event_handler->device->event_handler_lock, flags);
	1938	+ up_write(&event_handler->device->event_handler_rwsem);
785	1939	}
786	1940	EXPORT_SYMBOL(ib_register_event_handler);
787	1941
..	..	@@ -794,35 +1948,87 @@
794	1948	*/
795	1949	void ib_unregister_event_handler(struct ib_event_handler *event_handler)
796	1950	{
797		- unsigned long flags;
798		-
799		- spin_lock_irqsave(&event_handler->device->event_handler_lock, flags);
	1951	+ down_write(&event_handler->device->event_handler_rwsem);
800	1952	list_del(&event_handler->list);
801		- spin_unlock_irqrestore(&event_handler->device->event_handler_lock, flags);
	1953	+ up_write(&event_handler->device->event_handler_rwsem);
802	1954	}
803	1955	EXPORT_SYMBOL(ib_unregister_event_handler);
804	1956
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)
	1957	+void ib_dispatch_event_clients(struct ib_event *event)
814	1958	{
815		- unsigned long flags;
816	1959	struct ib_event_handler *handler;
817	1960
818		- spin_lock_irqsave(&event->device->event_handler_lock, flags);
	1961	+ down_read(&event->device->event_handler_rwsem);
819	1962
820	1963	list_for_each_entry(handler, &event->device->event_handler_list, list)
821	1964	handler->handler(handler, event);
822	1965
823		- spin_unlock_irqrestore(&event->device->event_handler_lock, flags);
	1966	+ up_read(&event->device->event_handler_rwsem);
824	1967	}
825		-EXPORT_SYMBOL(ib_dispatch_event);
	1968	+
	1969	+static int iw_query_port(struct ib_device *device,
	1970	+ u8 port_num,
	1971	+ struct ib_port_attr *port_attr)
	1972	+{
	1973	+ struct in_device *inetdev;
	1974	+ struct net_device *netdev;
	1975	+
	1976	+ memset(port_attr, 0, sizeof(*port_attr));
	1977	+
	1978	+ netdev = ib_device_get_netdev(device, port_num);
	1979	+ if (!netdev)
	1980	+ return -ENODEV;
	1981	+
	1982	+ port_attr->max_mtu = IB_MTU_4096;
	1983	+ port_attr->active_mtu = ib_mtu_int_to_enum(netdev->mtu);
	1984	+
	1985	+ if (!netif_carrier_ok(netdev)) {
	1986	+ port_attr->state = IB_PORT_DOWN;
	1987	+ port_attr->phys_state = IB_PORT_PHYS_STATE_DISABLED;
	1988	+ } else {
	1989	+ rcu_read_lock();
	1990	+ inetdev = __in_dev_get_rcu(netdev);
	1991	+
	1992	+ if (inetdev && inetdev->ifa_list) {
	1993	+ port_attr->state = IB_PORT_ACTIVE;
	1994	+ port_attr->phys_state = IB_PORT_PHYS_STATE_LINK_UP;
	1995	+ } else {
	1996	+ port_attr->state = IB_PORT_INIT;
	1997	+ port_attr->phys_state =
	1998	+ IB_PORT_PHYS_STATE_PORT_CONFIGURATION_TRAINING;
	1999	+ }
	2000	+
	2001	+ rcu_read_unlock();
	2002	+ }
	2003	+
	2004	+ dev_put(netdev);
	2005	+ return device->ops.query_port(device, port_num, port_attr);
	2006	+}
	2007	+
	2008	+static int __ib_query_port(struct ib_device *device,
	2009	+ u8 port_num,
	2010	+ struct ib_port_attr *port_attr)
	2011	+{
	2012	+ union ib_gid gid = {};
	2013	+ int err;
	2014	+
	2015	+ memset(port_attr, 0, sizeof(*port_attr));
	2016	+
	2017	+ err = device->ops.query_port(device, port_num, port_attr);
	2018	+ if (err \|\| port_attr->subnet_prefix)
	2019	+ return err;
	2020	+
	2021	+ if (rdma_port_get_link_layer(device, port_num) !=
	2022	+ IB_LINK_LAYER_INFINIBAND)
	2023	+ return 0;
	2024	+
	2025	+ err = device->ops.query_gid(device, port_num, 0, &gid);
	2026	+ if (err)
	2027	+ return err;
	2028	+
	2029	+ port_attr->subnet_prefix = be64_to_cpu(gid.global.subnet_prefix);
	2030	+ return 0;
	2031	+}
826	2032
827	2033	/**
828	2034	* ib_query_port - Query IB port attributes
..	..	@@ -837,28 +2043,197 @@
837	2043	u8 port_num,
838	2044	struct ib_port_attr *port_attr)
839	2045	{
840		- union ib_gid gid;
841		- int err;
842		-
843	2046	if (!rdma_is_port_valid(device, port_num))
844	2047	return -EINVAL;
845	2048
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;
	2049	+ if (rdma_protocol_iwarp(device, port_num))
	2050	+ return iw_query_port(device, port_num, port_attr);
	2051	+ else
	2052	+ return __ib_query_port(device, port_num, port_attr);
860	2053	}
861	2054	EXPORT_SYMBOL(ib_query_port);
	2055	+
	2056	+static void add_ndev_hash(struct ib_port_data *pdata)
	2057	+{
	2058	+ unsigned long flags;
	2059	+
	2060	+ might_sleep();
	2061	+
	2062	+ spin_lock_irqsave(&ndev_hash_lock, flags);
	2063	+ if (hash_hashed(&pdata->ndev_hash_link)) {
	2064	+ hash_del_rcu(&pdata->ndev_hash_link);
	2065	+ spin_unlock_irqrestore(&ndev_hash_lock, flags);
	2066	+ /*
	2067	+ * We cannot do hash_add_rcu after a hash_del_rcu until the
	2068	+ * grace period
	2069	+ */
	2070	+ synchronize_rcu();
	2071	+ spin_lock_irqsave(&ndev_hash_lock, flags);
	2072	+ }
	2073	+ if (pdata->netdev)
	2074	+ hash_add_rcu(ndev_hash, &pdata->ndev_hash_link,
	2075	+ (uintptr_t)pdata->netdev);
	2076	+ spin_unlock_irqrestore(&ndev_hash_lock, flags);
	2077	+}
	2078	+
	2079	+/**
	2080	+ * ib_device_set_netdev - Associate the ib_dev with an underlying net_device
	2081	+ * @ib_dev: Device to modify
	2082	+ * @ndev: net_device to affiliate, may be NULL
	2083	+ * @port: IB port the net_device is connected to
	2084	+ *
	2085	+ * Drivers should use this to link the ib_device to a netdev so the netdev
	2086	+ * shows up in interfaces like ib_enum_roce_netdev. Only one netdev may be
	2087	+ * affiliated with any port.
	2088	+ *
	2089	+ * The caller must ensure that the given ndev is not unregistered or
	2090	+ * unregistering, and that either the ib_device is unregistered or
	2091	+ * ib_device_set_netdev() is called with NULL when the ndev sends a
	2092	+ * NETDEV_UNREGISTER event.
	2093	+ */
	2094	+int ib_device_set_netdev(struct ib_device ib_dev, struct net_device ndev,
	2095	+ unsigned int port)
	2096	+{
	2097	+ struct net_device *old_ndev;
	2098	+ struct ib_port_data *pdata;
	2099	+ unsigned long flags;
	2100	+ int ret;
	2101	+
	2102	+ /*
	2103	+ * Drivers wish to call this before ib_register_driver, so we have to
	2104	+ * setup the port data early.
	2105	+ */
	2106	+ ret = alloc_port_data(ib_dev);
	2107	+ if (ret)
	2108	+ return ret;
	2109	+
	2110	+ if (!rdma_is_port_valid(ib_dev, port))
	2111	+ return -EINVAL;
	2112	+
	2113	+ pdata = &ib_dev->port_data[port];
	2114	+ spin_lock_irqsave(&pdata->netdev_lock, flags);
	2115	+ old_ndev = rcu_dereference_protected(
	2116	+ pdata->netdev, lockdep_is_held(&pdata->netdev_lock));
	2117	+ if (old_ndev == ndev) {
	2118	+ spin_unlock_irqrestore(&pdata->netdev_lock, flags);
	2119	+ return 0;
	2120	+ }
	2121	+
	2122	+ if (ndev)
	2123	+ dev_hold(ndev);
	2124	+ rcu_assign_pointer(pdata->netdev, ndev);
	2125	+ spin_unlock_irqrestore(&pdata->netdev_lock, flags);
	2126	+
	2127	+ add_ndev_hash(pdata);
	2128	+ if (old_ndev)
	2129	+ dev_put(old_ndev);
	2130	+
	2131	+ return 0;
	2132	+}
	2133	+EXPORT_SYMBOL(ib_device_set_netdev);
	2134	+
	2135	+static void free_netdevs(struct ib_device *ib_dev)
	2136	+{
	2137	+ unsigned long flags;
	2138	+ unsigned int port;
	2139	+
	2140	+ if (!ib_dev->port_data)
	2141	+ return;
	2142	+
	2143	+ rdma_for_each_port (ib_dev, port) {
	2144	+ struct ib_port_data *pdata = &ib_dev->port_data[port];
	2145	+ struct net_device *ndev;
	2146	+
	2147	+ spin_lock_irqsave(&pdata->netdev_lock, flags);
	2148	+ ndev = rcu_dereference_protected(
	2149	+ pdata->netdev, lockdep_is_held(&pdata->netdev_lock));
	2150	+ if (ndev) {
	2151	+ spin_lock(&ndev_hash_lock);
	2152	+ hash_del_rcu(&pdata->ndev_hash_link);
	2153	+ spin_unlock(&ndev_hash_lock);
	2154	+
	2155	+ /*
	2156	+ * If this is the last dev_put there is still a
	2157	+ * synchronize_rcu before the netdev is kfreed, so we
	2158	+ * can continue to rely on unlocked pointer
	2159	+ * comparisons after the put
	2160	+ */
	2161	+ rcu_assign_pointer(pdata->netdev, NULL);
	2162	+ dev_put(ndev);
	2163	+ }
	2164	+ spin_unlock_irqrestore(&pdata->netdev_lock, flags);
	2165	+ }
	2166	+}
	2167	+
	2168	+struct net_device ib_device_get_netdev(struct ib_device ib_dev,
	2169	+ unsigned int port)
	2170	+{
	2171	+ struct ib_port_data *pdata;
	2172	+ struct net_device *res;
	2173	+
	2174	+ if (!rdma_is_port_valid(ib_dev, port))
	2175	+ return NULL;
	2176	+
	2177	+ pdata = &ib_dev->port_data[port];
	2178	+
	2179	+ /*
	2180	+ * New drivers should use ib_device_set_netdev() not the legacy
	2181	+ * get_netdev().
	2182	+ */
	2183	+ if (ib_dev->ops.get_netdev)
	2184	+ res = ib_dev->ops.get_netdev(ib_dev, port);
	2185	+ else {
	2186	+ spin_lock(&pdata->netdev_lock);
	2187	+ res = rcu_dereference_protected(
	2188	+ pdata->netdev, lockdep_is_held(&pdata->netdev_lock));
	2189	+ if (res)
	2190	+ dev_hold(res);
	2191	+ spin_unlock(&pdata->netdev_lock);
	2192	+ }
	2193	+
	2194	+ /*
	2195	+ * If we are starting to unregister expedite things by preventing
	2196	+ * propagation of an unregistering netdev.
	2197	+ */
	2198	+ if (res && res->reg_state != NETREG_REGISTERED) {
	2199	+ dev_put(res);
	2200	+ return NULL;
	2201	+ }
	2202	+
	2203	+ return res;
	2204	+}
	2205	+
	2206	+/**
	2207	+ * ib_device_get_by_netdev - Find an IB device associated with a netdev
	2208	+ * @ndev: netdev to locate
	2209	+ * @driver_id: The driver ID that must match (RDMA_DRIVER_UNKNOWN matches all)
	2210	+ *
	2211	+ * Find and hold an ib_device that is associated with a netdev via
	2212	+ * ib_device_set_netdev(). The caller must call ib_device_put() on the
	2213	+ * returned pointer.
	2214	+ */
	2215	+struct ib_device ib_device_get_by_netdev(struct net_device ndev,
	2216	+ enum rdma_driver_id driver_id)
	2217	+{
	2218	+ struct ib_device *res = NULL;
	2219	+ struct ib_port_data *cur;
	2220	+
	2221	+ rcu_read_lock();
	2222	+ hash_for_each_possible_rcu (ndev_hash, cur, ndev_hash_link,
	2223	+ (uintptr_t)ndev) {
	2224	+ if (rcu_access_pointer(cur->netdev) == ndev &&
	2225	+ (driver_id == RDMA_DRIVER_UNKNOWN \|\|
	2226	+ cur->ib_dev->ops.driver_id == driver_id) &&
	2227	+ ib_device_try_get(cur->ib_dev)) {
	2228	+ res = cur->ib_dev;
	2229	+ break;
	2230	+ }
	2231	+ }
	2232	+ rcu_read_unlock();
	2233	+
	2234	+ return res;
	2235	+}
	2236	+EXPORT_SYMBOL(ib_device_get_by_netdev);
862	2237
863	2238	/**
864	2239	* ib_enum_roce_netdev - enumerate all RoCE ports
..	..	@@ -878,21 +2253,12 @@
878	2253	roce_netdev_callback cb,
879	2254	void *cookie)
880	2255	{
881		- u8 port;
	2256	+ unsigned int port;
882	2257
883		- for (port = rdma_start_port(ib_dev); port <= rdma_end_port(ib_dev);
884		- port++)
	2258	+ rdma_for_each_port (ib_dev, port)
885	2259	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		- }
	2260	+ struct net_device *idev =
	2261	+ ib_device_get_netdev(ib_dev, port);
896	2262
897	2263	if (filter(ib_dev, port, idev, filter_cookie))
898	2264	cb(ib_dev, port, idev, cookie);
..	..	@@ -919,11 +2285,12 @@
919	2285	void *cookie)
920	2286	{
921	2287	struct ib_device *dev;
	2288	+ unsigned long index;
922	2289
923		- down_read(&lists_rwsem);
924		- list_for_each_entry(dev, &device_list, core_list)
	2290	+ down_read(&devices_rwsem);
	2291	+ xa_for_each_marked (&devices, index, dev, DEVICE_REGISTERED)
925	2292	ib_enum_roce_netdev(dev, filter, filter_cookie, cb, cookie);
926		- up_read(&lists_rwsem);
	2293	+ up_read(&devices_rwsem);
927	2294	}
928	2295
929	2296	/**
..	..	@@ -935,19 +2302,22 @@
935	2302	int ib_enum_all_devs(nldev_callback nldev_cb, struct sk_buff *skb,
936	2303	struct netlink_callback *cb)
937	2304	{
	2305	+ unsigned long index;
938	2306	struct ib_device *dev;
939	2307	unsigned int idx = 0;
940	2308	int ret = 0;
941	2309
942		- down_read(&lists_rwsem);
943		- list_for_each_entry(dev, &device_list, core_list) {
	2310	+ down_read(&devices_rwsem);
	2311	+ xa_for_each_marked (&devices, index, dev, DEVICE_REGISTERED) {
	2312	+ if (!rdma_dev_access_netns(dev, sock_net(skb->sk)))
	2313	+ continue;
	2314	+
944	2315	ret = nldev_cb(dev, skb, cb, idx);
945	2316	if (ret)
946	2317	break;
947	2318	idx++;
948	2319	}
949		-
950		- up_read(&lists_rwsem);
	2320	+ up_read(&devices_rwsem);
951	2321	return ret;
952	2322	}
953	2323
..	..	@@ -963,7 +2333,13 @@
963	2333	int ib_query_pkey(struct ib_device *device,
964	2334	u8 port_num, u16 index, u16 *pkey)
965	2335	{
966		- return device->query_pkey(device, port_num, index, pkey);
	2336	+ if (!rdma_is_port_valid(device, port_num))
	2337	+ return -EINVAL;
	2338	+
	2339	+ if (!device->ops.query_pkey)
	2340	+ return -EOPNOTSUPP;
	2341	+
	2342	+ return device->ops.query_pkey(device, port_num, index, pkey);
967	2343	}
968	2344	EXPORT_SYMBOL(ib_query_pkey);
969	2345
..	..	@@ -980,11 +2356,11 @@
980	2356	int device_modify_mask,
981	2357	struct ib_device_modify *device_modify)
982	2358	{
983		- if (!device->modify_device)
984		- return -ENOSYS;
	2359	+ if (!device->ops.modify_device)
	2360	+ return -EOPNOTSUPP;
985	2361
986		- return device->modify_device(device, device_modify_mask,
987		- device_modify);
	2362	+ return device->ops.modify_device(device, device_modify_mask,
	2363	+ device_modify);
988	2364	}
989	2365	EXPORT_SYMBOL(ib_modify_device);
990	2366
..	..	@@ -1008,11 +2384,16 @@
1008	2384	if (!rdma_is_port_valid(device, port_num))
1009	2385	return -EINVAL;
1010	2386
1011		- if (device->modify_port)
1012		- rc = device->modify_port(device, port_num, port_modify_mask,
1013		- port_modify);
	2387	+ if (device->ops.modify_port)
	2388	+ rc = device->ops.modify_port(device, port_num,
	2389	+ port_modify_mask,
	2390	+ port_modify);
	2391	+ else if (rdma_protocol_roce(device, port_num) &&
	2392	+ ((port_modify->set_port_cap_mask & ~IB_PORT_CM_SUP) == 0 \|\|
	2393	+ (port_modify->clr_port_cap_mask & ~IB_PORT_CM_SUP) == 0))
	2394	+ rc = 0;
1014	2395	else
1015		- rc = rdma_protocol_roce(device, port_num) ? 0 : -ENOSYS;
	2396	+ rc = -EOPNOTSUPP;
1016	2397	return rc;
1017	2398	}
1018	2399	EXPORT_SYMBOL(ib_modify_port);
..	..	@@ -1030,13 +2411,15 @@
1030	2411	u8 port_num, u16 index)
1031	2412	{
1032	2413	union ib_gid tmp_gid;
1033		- int ret, port, i;
	2414	+ unsigned int port;
	2415	+ int ret, i;
1034	2416
1035		- for (port = rdma_start_port(device); port <= rdma_end_port(device); ++port) {
	2417	+ rdma_for_each_port (device, port) {
1036	2418	if (!rdma_protocol_ib(device, port))
1037	2419	continue;
1038	2420
1039		- for (i = 0; i < device->port_immutable[port].gid_tbl_len; ++i) {
	2421	+ for (i = 0; i < device->port_data[port].immutable.gid_tbl_len;
	2422	+ ++i) {
1040	2423	ret = rdma_query_gid(device, port, i, &tmp_gid);
1041	2424	if (ret)
1042	2425	continue;
..	..	@@ -1069,7 +2452,8 @@
1069	2452	u16 tmp_pkey;
1070	2453	int partial_ix = -1;
1071	2454
1072		- for (i = 0; i < device->port_immutable[port_num].pkey_tbl_len; ++i) {
	2455	+ for (i = 0; i < device->port_data[port_num].immutable.pkey_tbl_len;
	2456	+ ++i) {
1073	2457	ret = ib_query_pkey(device, port_num, i, &tmp_pkey);
1074	2458	if (ret)
1075	2459	return ret;
..	..	@@ -1102,6 +2486,7 @@
1102	2486	* @gid: A GID that the net_dev uses to communicate.
1103	2487	* @addr: Contains the IP address that the request specified as its
1104	2488	* destination.
	2489	+ *
1105	2490	*/
1106	2491	struct net_device ib_get_net_dev_by_params(struct ib_device dev,
1107	2492	u8 port,
..	..	@@ -1110,33 +2495,201 @@
1110	2495	const struct sockaddr *addr)
1111	2496	{
1112	2497	struct net_device *net_dev = NULL;
1113		- struct ib_client_data *context;
	2498	+ unsigned long index;
	2499	+ void *client_data;
1114	2500
1115	2501	if (!rdma_protocol_ib(dev, port))
1116	2502	return NULL;
1117	2503
1118		- down_read(&lists_rwsem);
	2504	+ /*
	2505	+ * Holding the read side guarantees that the client will not become
	2506	+ * unregistered while we are calling get_net_dev_by_params()
	2507	+ */
	2508	+ down_read(&dev->client_data_rwsem);
	2509	+ xan_for_each_marked (&dev->client_data, index, client_data,
	2510	+ CLIENT_DATA_REGISTERED) {
	2511	+ struct ib_client *client = xa_load(&clients, index);
1119	2512
1120		- list_for_each_entry(context, &dev->client_data_list, list) {
1121		- struct ib_client *client = context->client;
1122		-
1123		- if (context->going_down)
	2513	+ if (!client \|\| !client->get_net_dev_by_params)
1124	2514	continue;
1125	2515
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		- }
	2516	+ net_dev = client->get_net_dev_by_params(dev, port, pkey, gid,
	2517	+ addr, client_data);
	2518	+ if (net_dev)
	2519	+ break;
1133	2520	}
1134		-
1135		- up_read(&lists_rwsem);
	2521	+ up_read(&dev->client_data_rwsem);
1136	2522
1137	2523	return net_dev;
1138	2524	}
1139	2525	EXPORT_SYMBOL(ib_get_net_dev_by_params);
	2526	+
	2527	+void ib_set_device_ops(struct ib_device dev, const struct ib_device_ops ops)
	2528	+{
	2529	+ struct ib_device_ops *dev_ops = &dev->ops;
	2530	+#define SET_DEVICE_OP(ptr, name) \
	2531	+ do { \
	2532	+ if (ops->name) \
	2533	+ if (!((ptr)->name)) \
	2534	+ (ptr)->name = ops->name; \
	2535	+ } while (0)
	2536	+
	2537	+#define SET_OBJ_SIZE(ptr, name) SET_DEVICE_OP(ptr, size_##name)
	2538	+
	2539	+ if (ops->driver_id != RDMA_DRIVER_UNKNOWN) {
	2540	+ WARN_ON(dev_ops->driver_id != RDMA_DRIVER_UNKNOWN &&
	2541	+ dev_ops->driver_id != ops->driver_id);
	2542	+ dev_ops->driver_id = ops->driver_id;
	2543	+ }
	2544	+ if (ops->owner) {
	2545	+ WARN_ON(dev_ops->owner && dev_ops->owner != ops->owner);
	2546	+ dev_ops->owner = ops->owner;
	2547	+ }
	2548	+ if (ops->uverbs_abi_ver)
	2549	+ dev_ops->uverbs_abi_ver = ops->uverbs_abi_ver;
	2550	+
	2551	+ dev_ops->uverbs_no_driver_id_binding \|=
	2552	+ ops->uverbs_no_driver_id_binding;
	2553	+
	2554	+ SET_DEVICE_OP(dev_ops, add_gid);
	2555	+ SET_DEVICE_OP(dev_ops, advise_mr);
	2556	+ SET_DEVICE_OP(dev_ops, alloc_dm);
	2557	+ SET_DEVICE_OP(dev_ops, alloc_hw_stats);
	2558	+ SET_DEVICE_OP(dev_ops, alloc_mr);
	2559	+ SET_DEVICE_OP(dev_ops, alloc_mr_integrity);
	2560	+ SET_DEVICE_OP(dev_ops, alloc_mw);
	2561	+ SET_DEVICE_OP(dev_ops, alloc_pd);
	2562	+ SET_DEVICE_OP(dev_ops, alloc_rdma_netdev);
	2563	+ SET_DEVICE_OP(dev_ops, alloc_ucontext);
	2564	+ SET_DEVICE_OP(dev_ops, alloc_xrcd);
	2565	+ SET_DEVICE_OP(dev_ops, attach_mcast);
	2566	+ SET_DEVICE_OP(dev_ops, check_mr_status);
	2567	+ SET_DEVICE_OP(dev_ops, counter_alloc_stats);
	2568	+ SET_DEVICE_OP(dev_ops, counter_bind_qp);
	2569	+ SET_DEVICE_OP(dev_ops, counter_dealloc);
	2570	+ SET_DEVICE_OP(dev_ops, counter_unbind_qp);
	2571	+ SET_DEVICE_OP(dev_ops, counter_update_stats);
	2572	+ SET_DEVICE_OP(dev_ops, create_ah);
	2573	+ SET_DEVICE_OP(dev_ops, create_counters);
	2574	+ SET_DEVICE_OP(dev_ops, create_cq);
	2575	+ SET_DEVICE_OP(dev_ops, create_flow);
	2576	+ SET_DEVICE_OP(dev_ops, create_flow_action_esp);
	2577	+ SET_DEVICE_OP(dev_ops, create_qp);
	2578	+ SET_DEVICE_OP(dev_ops, create_rwq_ind_table);
	2579	+ SET_DEVICE_OP(dev_ops, create_srq);
	2580	+ SET_DEVICE_OP(dev_ops, create_wq);
	2581	+ SET_DEVICE_OP(dev_ops, dealloc_dm);
	2582	+ SET_DEVICE_OP(dev_ops, dealloc_driver);
	2583	+ SET_DEVICE_OP(dev_ops, dealloc_mw);
	2584	+ SET_DEVICE_OP(dev_ops, dealloc_pd);
	2585	+ SET_DEVICE_OP(dev_ops, dealloc_ucontext);
	2586	+ SET_DEVICE_OP(dev_ops, dealloc_xrcd);
	2587	+ SET_DEVICE_OP(dev_ops, del_gid);
	2588	+ SET_DEVICE_OP(dev_ops, dereg_mr);
	2589	+ SET_DEVICE_OP(dev_ops, destroy_ah);
	2590	+ SET_DEVICE_OP(dev_ops, destroy_counters);
	2591	+ SET_DEVICE_OP(dev_ops, destroy_cq);
	2592	+ SET_DEVICE_OP(dev_ops, destroy_flow);
	2593	+ SET_DEVICE_OP(dev_ops, destroy_flow_action);
	2594	+ SET_DEVICE_OP(dev_ops, destroy_qp);
	2595	+ SET_DEVICE_OP(dev_ops, destroy_rwq_ind_table);
	2596	+ SET_DEVICE_OP(dev_ops, destroy_srq);
	2597	+ SET_DEVICE_OP(dev_ops, destroy_wq);
	2598	+ SET_DEVICE_OP(dev_ops, detach_mcast);
	2599	+ SET_DEVICE_OP(dev_ops, disassociate_ucontext);
	2600	+ SET_DEVICE_OP(dev_ops, drain_rq);
	2601	+ SET_DEVICE_OP(dev_ops, drain_sq);
	2602	+ SET_DEVICE_OP(dev_ops, enable_driver);
	2603	+ SET_DEVICE_OP(dev_ops, fill_res_cm_id_entry);
	2604	+ SET_DEVICE_OP(dev_ops, fill_res_cq_entry);
	2605	+ SET_DEVICE_OP(dev_ops, fill_res_cq_entry_raw);
	2606	+ SET_DEVICE_OP(dev_ops, fill_res_mr_entry);
	2607	+ SET_DEVICE_OP(dev_ops, fill_res_mr_entry_raw);
	2608	+ SET_DEVICE_OP(dev_ops, fill_res_qp_entry);
	2609	+ SET_DEVICE_OP(dev_ops, fill_res_qp_entry_raw);
	2610	+ SET_DEVICE_OP(dev_ops, fill_stat_mr_entry);
	2611	+ SET_DEVICE_OP(dev_ops, get_dev_fw_str);
	2612	+ SET_DEVICE_OP(dev_ops, get_dma_mr);
	2613	+ SET_DEVICE_OP(dev_ops, get_hw_stats);
	2614	+ SET_DEVICE_OP(dev_ops, get_link_layer);
	2615	+ SET_DEVICE_OP(dev_ops, get_netdev);
	2616	+ SET_DEVICE_OP(dev_ops, get_port_immutable);
	2617	+ SET_DEVICE_OP(dev_ops, get_vector_affinity);
	2618	+ SET_DEVICE_OP(dev_ops, get_vf_config);
	2619	+ SET_DEVICE_OP(dev_ops, get_vf_guid);
	2620	+ SET_DEVICE_OP(dev_ops, get_vf_stats);
	2621	+ SET_DEVICE_OP(dev_ops, init_port);
	2622	+ SET_DEVICE_OP(dev_ops, iw_accept);
	2623	+ SET_DEVICE_OP(dev_ops, iw_add_ref);
	2624	+ SET_DEVICE_OP(dev_ops, iw_connect);
	2625	+ SET_DEVICE_OP(dev_ops, iw_create_listen);
	2626	+ SET_DEVICE_OP(dev_ops, iw_destroy_listen);
	2627	+ SET_DEVICE_OP(dev_ops, iw_get_qp);
	2628	+ SET_DEVICE_OP(dev_ops, iw_reject);
	2629	+ SET_DEVICE_OP(dev_ops, iw_rem_ref);
	2630	+ SET_DEVICE_OP(dev_ops, map_mr_sg);
	2631	+ SET_DEVICE_OP(dev_ops, map_mr_sg_pi);
	2632	+ SET_DEVICE_OP(dev_ops, mmap);
	2633	+ SET_DEVICE_OP(dev_ops, mmap_free);
	2634	+ SET_DEVICE_OP(dev_ops, modify_ah);
	2635	+ SET_DEVICE_OP(dev_ops, modify_cq);
	2636	+ SET_DEVICE_OP(dev_ops, modify_device);
	2637	+ SET_DEVICE_OP(dev_ops, modify_flow_action_esp);
	2638	+ SET_DEVICE_OP(dev_ops, modify_port);
	2639	+ SET_DEVICE_OP(dev_ops, modify_qp);
	2640	+ SET_DEVICE_OP(dev_ops, modify_srq);
	2641	+ SET_DEVICE_OP(dev_ops, modify_wq);
	2642	+ SET_DEVICE_OP(dev_ops, peek_cq);
	2643	+ SET_DEVICE_OP(dev_ops, poll_cq);
	2644	+ SET_DEVICE_OP(dev_ops, post_recv);
	2645	+ SET_DEVICE_OP(dev_ops, post_send);
	2646	+ SET_DEVICE_OP(dev_ops, post_srq_recv);
	2647	+ SET_DEVICE_OP(dev_ops, process_mad);
	2648	+ SET_DEVICE_OP(dev_ops, query_ah);
	2649	+ SET_DEVICE_OP(dev_ops, query_device);
	2650	+ SET_DEVICE_OP(dev_ops, query_gid);
	2651	+ SET_DEVICE_OP(dev_ops, query_pkey);
	2652	+ SET_DEVICE_OP(dev_ops, query_port);
	2653	+ SET_DEVICE_OP(dev_ops, query_qp);
	2654	+ SET_DEVICE_OP(dev_ops, query_srq);
	2655	+ SET_DEVICE_OP(dev_ops, query_ucontext);
	2656	+ SET_DEVICE_OP(dev_ops, rdma_netdev_get_params);
	2657	+ SET_DEVICE_OP(dev_ops, read_counters);
	2658	+ SET_DEVICE_OP(dev_ops, reg_dm_mr);
	2659	+ SET_DEVICE_OP(dev_ops, reg_user_mr);
	2660	+ SET_DEVICE_OP(dev_ops, req_ncomp_notif);
	2661	+ SET_DEVICE_OP(dev_ops, req_notify_cq);
	2662	+ SET_DEVICE_OP(dev_ops, rereg_user_mr);
	2663	+ SET_DEVICE_OP(dev_ops, resize_cq);
	2664	+ SET_DEVICE_OP(dev_ops, set_vf_guid);
	2665	+ SET_DEVICE_OP(dev_ops, set_vf_link_state);
	2666	+
	2667	+ SET_OBJ_SIZE(dev_ops, ib_ah);
	2668	+ SET_OBJ_SIZE(dev_ops, ib_counters);
	2669	+ SET_OBJ_SIZE(dev_ops, ib_cq);
	2670	+ SET_OBJ_SIZE(dev_ops, ib_mw);
	2671	+ SET_OBJ_SIZE(dev_ops, ib_pd);
	2672	+ SET_OBJ_SIZE(dev_ops, ib_rwq_ind_table);
	2673	+ SET_OBJ_SIZE(dev_ops, ib_srq);
	2674	+ SET_OBJ_SIZE(dev_ops, ib_ucontext);
	2675	+ SET_OBJ_SIZE(dev_ops, ib_xrcd);
	2676	+}
	2677	+EXPORT_SYMBOL(ib_set_device_ops);
	2678	+
	2679	+#ifdef CONFIG_INFINIBAND_VIRT_DMA
	2680	+int ib_dma_virt_map_sg(struct ib_device dev, struct scatterlist sg, int nents)
	2681	+{
	2682	+ struct scatterlist *s;
	2683	+ int i;
	2684	+
	2685	+ for_each_sg(sg, s, nents, i) {
	2686	+ sg_dma_address(s) = (uintptr_t)sg_virt(s);
	2687	+ sg_dma_len(s) = s->length;
	2688	+ }
	2689	+ return nents;
	2690	+}
	2691	+EXPORT_SYMBOL(ib_dma_virt_map_sg);
	2692	+#endif /* CONFIG_INFINIBAND_VIRT_DMA */
1140	2693
1141	2694	static const struct rdma_nl_cbs ibnl_ls_cb_table[RDMA_NL_LS_NUM_OPS] = {
1142	2695	[RDMA_NL_LS_OP_RESOLVE] = {
..	..	@@ -1183,15 +2736,11 @@
1183	2736	goto err_comp_unbound;
1184	2737	}
1185	2738
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		- }
	2739	+ rdma_nl_init();
1191	2740
1192	2741	ret = addr_init();
1193	2742	if (ret) {
1194		- pr_warn("Could't init IB address resolution\n");
	2743	+ pr_warn("Couldn't init IB address resolution\n");
1195	2744	goto err_ibnl;
1196	2745	}
1197	2746
..	..	@@ -1207,18 +2756,34 @@
1207	2756	goto err_mad;
1208	2757	}
1209	2758
1210		- ret = register_lsm_notifier(&ibdev_lsm_nb);
	2759	+ ret = register_blocking_lsm_notifier(&ibdev_lsm_nb);
1211	2760	if (ret) {
1212	2761	pr_warn("Couldn't register LSM notifier. ret %d\n", ret);
1213	2762	goto err_sa;
1214	2763	}
1215	2764
	2765	+ ret = register_pernet_device(&rdma_dev_net_ops);
	2766	+ if (ret) {
	2767	+ pr_warn("Couldn't init compat dev. ret %d\n", ret);
	2768	+ goto err_compat;
	2769	+ }
	2770	+
1216	2771	nldev_init();
1217	2772	rdma_nl_register(RDMA_NL_LS, ibnl_ls_cb_table);
1218		- roce_gid_mgmt_init();
	2773	+ ret = roce_gid_mgmt_init();
	2774	+ if (ret) {
	2775	+ pr_warn("Couldn't init RoCE GID management\n");
	2776	+ goto err_parent;
	2777	+ }
1219	2778
1220	2779	return 0;
1221	2780
	2781	+err_parent:
	2782	+ rdma_nl_unregister(RDMA_NL_LS);
	2783	+ nldev_exit();
	2784	+ unregister_pernet_device(&rdma_dev_net_ops);
	2785	+err_compat:
	2786	+ unregister_blocking_lsm_notifier(&ibdev_lsm_nb);
1222	2787	err_sa:
1223	2788	ib_sa_cleanup();
1224	2789	err_mad:
..	..	@@ -1226,8 +2791,6 @@
1226	2791	err_addr:
1227	2792	addr_cleanup();
1228	2793	err_ibnl:
1229		- rdma_nl_exit();
1230		-err_sysfs:
1231	2794	class_unregister(&ib_class);
1232	2795	err_comp_unbound:
1233	2796	destroy_workqueue(ib_comp_unbound_wq);
..	..	@@ -1241,9 +2804,10 @@
1241	2804	static void __exit ib_core_cleanup(void)
1242	2805	{
1243	2806	roce_gid_mgmt_cleanup();
1244		- nldev_exit();
1245	2807	rdma_nl_unregister(RDMA_NL_LS);
1246		- unregister_lsm_notifier(&ibdev_lsm_nb);
	2808	+ nldev_exit();
	2809	+ unregister_pernet_device(&rdma_dev_net_ops);
	2810	+ unregister_blocking_lsm_notifier(&ibdev_lsm_nb);
1247	2811	ib_sa_cleanup();
1248	2812	ib_mad_cleanup();
1249	2813	addr_cleanup();
..	..	@@ -1253,9 +2817,15 @@
1253	2817	destroy_workqueue(ib_comp_wq);
1254	2818	/* Make sure that any pending umem accounting work is done. */
1255	2819	destroy_workqueue(ib_wq);
	2820	+ flush_workqueue(system_unbound_wq);
	2821	+ WARN_ON(!xa_empty(&clients));
	2822	+ WARN_ON(!xa_empty(&devices));
1256	2823	}
1257	2824
1258	2825	MODULE_ALIAS_RDMA_NETLINK(RDMA_NL_LS, 4);
1259	2826
1260		-subsys_initcall(ib_core_init);
	2827	+/* ib core relies on netdev stack to first register net_ns_type_operations
	2828	+ * ns kobject type before ib_core initialization.
	2829	+ */
	2830	+fs_initcall(ib_core_init);
1261	2831	module_exit(ib_core_cleanup);