Loading...
1// SPDX-License-Identifier: GPL-2.0
2/* Copyright(c) 2016-2019 Intel Corporation. All rights reserved. */
3#include <linux/memremap.h>
4#include <linux/pagemap.h>
5#include <linux/memory.h>
6#include <linux/module.h>
7#include <linux/device.h>
8#include <linux/pfn_t.h>
9#include <linux/slab.h>
10#include <linux/dax.h>
11#include <linux/fs.h>
12#include <linux/mm.h>
13#include <linux/mman.h>
14#include <linux/memory-tiers.h>
15#include "dax-private.h"
16#include "bus.h"
17
18/*
19 * Default abstract distance assigned to the NUMA node onlined
20 * by DAX/kmem if the low level platform driver didn't initialize
21 * one for this NUMA node.
22 */
23#define MEMTIER_DEFAULT_DAX_ADISTANCE (MEMTIER_ADISTANCE_DRAM * 5)
24
25/* Memory resource name used for add_memory_driver_managed(). */
26static const char *kmem_name;
27/* Set if any memory will remain added when the driver will be unloaded. */
28static bool any_hotremove_failed;
29
30static int dax_kmem_range(struct dev_dax *dev_dax, int i, struct range *r)
31{
32 struct dev_dax_range *dax_range = &dev_dax->ranges[i];
33 struct range *range = &dax_range->range;
34
35 /* memory-block align the hotplug range */
36 r->start = ALIGN(range->start, memory_block_size_bytes());
37 r->end = ALIGN_DOWN(range->end + 1, memory_block_size_bytes()) - 1;
38 if (r->start >= r->end) {
39 r->start = range->start;
40 r->end = range->end;
41 return -ENOSPC;
42 }
43 return 0;
44}
45
46struct dax_kmem_data {
47 const char *res_name;
48 int mgid;
49 struct resource *res[];
50};
51
52static struct memory_dev_type *dax_slowmem_type;
53static int dev_dax_kmem_probe(struct dev_dax *dev_dax)
54{
55 struct device *dev = &dev_dax->dev;
56 unsigned long total_len = 0;
57 struct dax_kmem_data *data;
58 int i, rc, mapped = 0;
59 int numa_node;
60
61 /*
62 * Ensure good NUMA information for the persistent memory.
63 * Without this check, there is a risk that slow memory
64 * could be mixed in a node with faster memory, causing
65 * unavoidable performance issues.
66 */
67 numa_node = dev_dax->target_node;
68 if (numa_node < 0) {
69 dev_warn(dev, "rejecting DAX region with invalid node: %d\n",
70 numa_node);
71 return -EINVAL;
72 }
73
74 for (i = 0; i < dev_dax->nr_range; i++) {
75 struct range range;
76
77 rc = dax_kmem_range(dev_dax, i, &range);
78 if (rc) {
79 dev_info(dev, "mapping%d: %#llx-%#llx too small after alignment\n",
80 i, range.start, range.end);
81 continue;
82 }
83 total_len += range_len(&range);
84 }
85
86 if (!total_len) {
87 dev_warn(dev, "rejecting DAX region without any memory after alignment\n");
88 return -EINVAL;
89 }
90
91 init_node_memory_type(numa_node, dax_slowmem_type);
92
93 rc = -ENOMEM;
94 data = kzalloc(struct_size(data, res, dev_dax->nr_range), GFP_KERNEL);
95 if (!data)
96 goto err_dax_kmem_data;
97
98 data->res_name = kstrdup(dev_name(dev), GFP_KERNEL);
99 if (!data->res_name)
100 goto err_res_name;
101
102 rc = memory_group_register_static(numa_node, total_len);
103 if (rc < 0)
104 goto err_reg_mgid;
105 data->mgid = rc;
106
107 for (i = 0; i < dev_dax->nr_range; i++) {
108 struct resource *res;
109 struct range range;
110
111 rc = dax_kmem_range(dev_dax, i, &range);
112 if (rc)
113 continue;
114
115 /* Region is permanently reserved if hotremove fails. */
116 res = request_mem_region(range.start, range_len(&range), data->res_name);
117 if (!res) {
118 dev_warn(dev, "mapping%d: %#llx-%#llx could not reserve region\n",
119 i, range.start, range.end);
120 /*
121 * Once some memory has been onlined we can't
122 * assume that it can be un-onlined safely.
123 */
124 if (mapped)
125 continue;
126 rc = -EBUSY;
127 goto err_request_mem;
128 }
129 data->res[i] = res;
130
131 /*
132 * Set flags appropriate for System RAM. Leave ..._BUSY clear
133 * so that add_memory() can add a child resource. Do not
134 * inherit flags from the parent since it may set new flags
135 * unknown to us that will break add_memory() below.
136 */
137 res->flags = IORESOURCE_SYSTEM_RAM;
138
139 /*
140 * Ensure that future kexec'd kernels will not treat
141 * this as RAM automatically.
142 */
143 rc = add_memory_driver_managed(data->mgid, range.start,
144 range_len(&range), kmem_name, MHP_NID_IS_MGID);
145
146 if (rc) {
147 dev_warn(dev, "mapping%d: %#llx-%#llx memory add failed\n",
148 i, range.start, range.end);
149 release_resource(res);
150 kfree(res);
151 data->res[i] = NULL;
152 if (mapped)
153 continue;
154 goto err_request_mem;
155 }
156 mapped++;
157 }
158
159 dev_set_drvdata(dev, data);
160
161 return 0;
162
163err_request_mem:
164 memory_group_unregister(data->mgid);
165err_reg_mgid:
166 kfree(data->res_name);
167err_res_name:
168 kfree(data);
169err_dax_kmem_data:
170 clear_node_memory_type(numa_node, dax_slowmem_type);
171 return rc;
172}
173
174#ifdef CONFIG_MEMORY_HOTREMOVE
175static void dev_dax_kmem_remove(struct dev_dax *dev_dax)
176{
177 int i, success = 0;
178 int node = dev_dax->target_node;
179 struct device *dev = &dev_dax->dev;
180 struct dax_kmem_data *data = dev_get_drvdata(dev);
181
182 /*
183 * We have one shot for removing memory, if some memory blocks were not
184 * offline prior to calling this function remove_memory() will fail, and
185 * there is no way to hotremove this memory until reboot because device
186 * unbind will succeed even if we return failure.
187 */
188 for (i = 0; i < dev_dax->nr_range; i++) {
189 struct range range;
190 int rc;
191
192 rc = dax_kmem_range(dev_dax, i, &range);
193 if (rc)
194 continue;
195
196 rc = remove_memory(range.start, range_len(&range));
197 if (rc == 0) {
198 release_resource(data->res[i]);
199 kfree(data->res[i]);
200 data->res[i] = NULL;
201 success++;
202 continue;
203 }
204 any_hotremove_failed = true;
205 dev_err(dev,
206 "mapping%d: %#llx-%#llx cannot be hotremoved until the next reboot\n",
207 i, range.start, range.end);
208 }
209
210 if (success >= dev_dax->nr_range) {
211 memory_group_unregister(data->mgid);
212 kfree(data->res_name);
213 kfree(data);
214 dev_set_drvdata(dev, NULL);
215 /*
216 * Clear the memtype association on successful unplug.
217 * If not, we have memory blocks left which can be
218 * offlined/onlined later. We need to keep memory_dev_type
219 * for that. This implies this reference will be around
220 * till next reboot.
221 */
222 clear_node_memory_type(node, dax_slowmem_type);
223 }
224}
225#else
226static void dev_dax_kmem_remove(struct dev_dax *dev_dax)
227{
228 /*
229 * Without hotremove purposely leak the request_mem_region() for the
230 * device-dax range and return '0' to ->remove() attempts. The removal
231 * of the device from the driver always succeeds, but the region is
232 * permanently pinned as reserved by the unreleased
233 * request_mem_region().
234 */
235 any_hotremove_failed = true;
236}
237#endif /* CONFIG_MEMORY_HOTREMOVE */
238
239static struct dax_device_driver device_dax_kmem_driver = {
240 .probe = dev_dax_kmem_probe,
241 .remove = dev_dax_kmem_remove,
242};
243
244static int __init dax_kmem_init(void)
245{
246 int rc;
247
248 /* Resource name is permanently allocated if any hotremove fails. */
249 kmem_name = kstrdup_const("System RAM (kmem)", GFP_KERNEL);
250 if (!kmem_name)
251 return -ENOMEM;
252
253 dax_slowmem_type = alloc_memory_type(MEMTIER_DEFAULT_DAX_ADISTANCE);
254 if (IS_ERR(dax_slowmem_type)) {
255 rc = PTR_ERR(dax_slowmem_type);
256 goto err_dax_slowmem_type;
257 }
258
259 rc = dax_driver_register(&device_dax_kmem_driver);
260 if (rc)
261 goto error_dax_driver;
262
263 return rc;
264
265error_dax_driver:
266 destroy_memory_type(dax_slowmem_type);
267err_dax_slowmem_type:
268 kfree_const(kmem_name);
269 return rc;
270}
271
272static void __exit dax_kmem_exit(void)
273{
274 dax_driver_unregister(&device_dax_kmem_driver);
275 if (!any_hotremove_failed)
276 kfree_const(kmem_name);
277 destroy_memory_type(dax_slowmem_type);
278}
279
280MODULE_AUTHOR("Intel Corporation");
281MODULE_LICENSE("GPL v2");
282module_init(dax_kmem_init);
283module_exit(dax_kmem_exit);
284MODULE_ALIAS_DAX_DEVICE(0);
1// SPDX-License-Identifier: GPL-2.0
2/* Copyright(c) 2016-2019 Intel Corporation. All rights reserved. */
3#include <linux/memremap.h>
4#include <linux/pagemap.h>
5#include <linux/memory.h>
6#include <linux/module.h>
7#include <linux/device.h>
8#include <linux/pfn_t.h>
9#include <linux/slab.h>
10#include <linux/dax.h>
11#include <linux/fs.h>
12#include <linux/mm.h>
13#include <linux/mman.h>
14#include <linux/memory-tiers.h>
15#include <linux/memory_hotplug.h>
16#include "dax-private.h"
17#include "bus.h"
18
19/*
20 * Default abstract distance assigned to the NUMA node onlined
21 * by DAX/kmem if the low level platform driver didn't initialize
22 * one for this NUMA node.
23 */
24#define MEMTIER_DEFAULT_DAX_ADISTANCE (MEMTIER_ADISTANCE_DRAM * 5)
25
26/* Memory resource name used for add_memory_driver_managed(). */
27static const char *kmem_name;
28/* Set if any memory will remain added when the driver will be unloaded. */
29static bool any_hotremove_failed;
30
31static int dax_kmem_range(struct dev_dax *dev_dax, int i, struct range *r)
32{
33 struct dev_dax_range *dax_range = &dev_dax->ranges[i];
34 struct range *range = &dax_range->range;
35
36 /* memory-block align the hotplug range */
37 r->start = ALIGN(range->start, memory_block_size_bytes());
38 r->end = ALIGN_DOWN(range->end + 1, memory_block_size_bytes()) - 1;
39 if (r->start >= r->end) {
40 r->start = range->start;
41 r->end = range->end;
42 return -ENOSPC;
43 }
44 return 0;
45}
46
47struct dax_kmem_data {
48 const char *res_name;
49 int mgid;
50 struct resource *res[];
51};
52
53static DEFINE_MUTEX(kmem_memory_type_lock);
54static LIST_HEAD(kmem_memory_types);
55
56static struct memory_dev_type *kmem_find_alloc_memory_type(int adist)
57{
58 guard(mutex)(&kmem_memory_type_lock);
59 return mt_find_alloc_memory_type(adist, &kmem_memory_types);
60}
61
62static void kmem_put_memory_types(void)
63{
64 guard(mutex)(&kmem_memory_type_lock);
65 mt_put_memory_types(&kmem_memory_types);
66}
67
68static int dev_dax_kmem_probe(struct dev_dax *dev_dax)
69{
70 struct device *dev = &dev_dax->dev;
71 unsigned long total_len = 0;
72 struct dax_kmem_data *data;
73 struct memory_dev_type *mtype;
74 int i, rc, mapped = 0;
75 mhp_t mhp_flags;
76 int numa_node;
77 int adist = MEMTIER_DEFAULT_DAX_ADISTANCE;
78
79 /*
80 * Ensure good NUMA information for the persistent memory.
81 * Without this check, there is a risk that slow memory
82 * could be mixed in a node with faster memory, causing
83 * unavoidable performance issues.
84 */
85 numa_node = dev_dax->target_node;
86 if (numa_node < 0) {
87 dev_warn(dev, "rejecting DAX region with invalid node: %d\n",
88 numa_node);
89 return -EINVAL;
90 }
91
92 mt_calc_adistance(numa_node, &adist);
93 mtype = kmem_find_alloc_memory_type(adist);
94 if (IS_ERR(mtype))
95 return PTR_ERR(mtype);
96
97 for (i = 0; i < dev_dax->nr_range; i++) {
98 struct range range;
99
100 rc = dax_kmem_range(dev_dax, i, &range);
101 if (rc) {
102 dev_info(dev, "mapping%d: %#llx-%#llx too small after alignment\n",
103 i, range.start, range.end);
104 continue;
105 }
106 total_len += range_len(&range);
107 }
108
109 if (!total_len) {
110 dev_warn(dev, "rejecting DAX region without any memory after alignment\n");
111 return -EINVAL;
112 }
113
114 init_node_memory_type(numa_node, mtype);
115
116 rc = -ENOMEM;
117 data = kzalloc(struct_size(data, res, dev_dax->nr_range), GFP_KERNEL);
118 if (!data)
119 goto err_dax_kmem_data;
120
121 data->res_name = kstrdup(dev_name(dev), GFP_KERNEL);
122 if (!data->res_name)
123 goto err_res_name;
124
125 rc = memory_group_register_static(numa_node, PFN_UP(total_len));
126 if (rc < 0)
127 goto err_reg_mgid;
128 data->mgid = rc;
129
130 for (i = 0; i < dev_dax->nr_range; i++) {
131 struct resource *res;
132 struct range range;
133
134 rc = dax_kmem_range(dev_dax, i, &range);
135 if (rc)
136 continue;
137
138 /* Region is permanently reserved if hotremove fails. */
139 res = request_mem_region(range.start, range_len(&range), data->res_name);
140 if (!res) {
141 dev_warn(dev, "mapping%d: %#llx-%#llx could not reserve region\n",
142 i, range.start, range.end);
143 /*
144 * Once some memory has been onlined we can't
145 * assume that it can be un-onlined safely.
146 */
147 if (mapped)
148 continue;
149 rc = -EBUSY;
150 goto err_request_mem;
151 }
152 data->res[i] = res;
153
154 /*
155 * Set flags appropriate for System RAM. Leave ..._BUSY clear
156 * so that add_memory() can add a child resource. Do not
157 * inherit flags from the parent since it may set new flags
158 * unknown to us that will break add_memory() below.
159 */
160 res->flags = IORESOURCE_SYSTEM_RAM;
161
162 mhp_flags = MHP_NID_IS_MGID;
163 if (dev_dax->memmap_on_memory)
164 mhp_flags |= MHP_MEMMAP_ON_MEMORY;
165
166 /*
167 * Ensure that future kexec'd kernels will not treat
168 * this as RAM automatically.
169 */
170 rc = add_memory_driver_managed(data->mgid, range.start,
171 range_len(&range), kmem_name, mhp_flags);
172
173 if (rc) {
174 dev_warn(dev, "mapping%d: %#llx-%#llx memory add failed\n",
175 i, range.start, range.end);
176 remove_resource(res);
177 kfree(res);
178 data->res[i] = NULL;
179 if (mapped)
180 continue;
181 goto err_request_mem;
182 }
183 mapped++;
184 }
185
186 dev_set_drvdata(dev, data);
187
188 return 0;
189
190err_request_mem:
191 memory_group_unregister(data->mgid);
192err_reg_mgid:
193 kfree(data->res_name);
194err_res_name:
195 kfree(data);
196err_dax_kmem_data:
197 clear_node_memory_type(numa_node, mtype);
198 return rc;
199}
200
201#ifdef CONFIG_MEMORY_HOTREMOVE
202static void dev_dax_kmem_remove(struct dev_dax *dev_dax)
203{
204 int i, success = 0;
205 int node = dev_dax->target_node;
206 struct device *dev = &dev_dax->dev;
207 struct dax_kmem_data *data = dev_get_drvdata(dev);
208
209 /*
210 * We have one shot for removing memory, if some memory blocks were not
211 * offline prior to calling this function remove_memory() will fail, and
212 * there is no way to hotremove this memory until reboot because device
213 * unbind will succeed even if we return failure.
214 */
215 for (i = 0; i < dev_dax->nr_range; i++) {
216 struct range range;
217 int rc;
218
219 rc = dax_kmem_range(dev_dax, i, &range);
220 if (rc)
221 continue;
222
223 rc = remove_memory(range.start, range_len(&range));
224 if (rc == 0) {
225 remove_resource(data->res[i]);
226 kfree(data->res[i]);
227 data->res[i] = NULL;
228 success++;
229 continue;
230 }
231 any_hotremove_failed = true;
232 dev_err(dev,
233 "mapping%d: %#llx-%#llx cannot be hotremoved until the next reboot\n",
234 i, range.start, range.end);
235 }
236
237 if (success >= dev_dax->nr_range) {
238 memory_group_unregister(data->mgid);
239 kfree(data->res_name);
240 kfree(data);
241 dev_set_drvdata(dev, NULL);
242 /*
243 * Clear the memtype association on successful unplug.
244 * If not, we have memory blocks left which can be
245 * offlined/onlined later. We need to keep memory_dev_type
246 * for that. This implies this reference will be around
247 * till next reboot.
248 */
249 clear_node_memory_type(node, NULL);
250 }
251}
252#else
253static void dev_dax_kmem_remove(struct dev_dax *dev_dax)
254{
255 /*
256 * Without hotremove purposely leak the request_mem_region() for the
257 * device-dax range and return '0' to ->remove() attempts. The removal
258 * of the device from the driver always succeeds, but the region is
259 * permanently pinned as reserved by the unreleased
260 * request_mem_region().
261 */
262 any_hotremove_failed = true;
263}
264#endif /* CONFIG_MEMORY_HOTREMOVE */
265
266static struct dax_device_driver device_dax_kmem_driver = {
267 .probe = dev_dax_kmem_probe,
268 .remove = dev_dax_kmem_remove,
269 .type = DAXDRV_KMEM_TYPE,
270};
271
272static int __init dax_kmem_init(void)
273{
274 int rc;
275
276 /* Resource name is permanently allocated if any hotremove fails. */
277 kmem_name = kstrdup_const("System RAM (kmem)", GFP_KERNEL);
278 if (!kmem_name)
279 return -ENOMEM;
280
281 rc = dax_driver_register(&device_dax_kmem_driver);
282 if (rc)
283 goto error_dax_driver;
284
285 return rc;
286
287error_dax_driver:
288 kmem_put_memory_types();
289 kfree_const(kmem_name);
290 return rc;
291}
292
293static void __exit dax_kmem_exit(void)
294{
295 dax_driver_unregister(&device_dax_kmem_driver);
296 if (!any_hotremove_failed)
297 kfree_const(kmem_name);
298 kmem_put_memory_types();
299}
300
301MODULE_AUTHOR("Intel Corporation");
302MODULE_DESCRIPTION("KMEM DAX: map dax-devices as System-RAM");
303MODULE_LICENSE("GPL v2");
304module_init(dax_kmem_init);
305module_exit(dax_kmem_exit);
306MODULE_ALIAS_DAX_DEVICE(0);