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/slab.h>
9 #include <linux/dax.h>
10 #include <linux/fs.h>
11 #include <linux/mm.h>
12 #include <linux/mman.h>
13 #include <linux/memory-tiers.h>
14 #include <linux/memory_hotplug.h>
15 #include <linux/string_helpers.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(). */
27 static const char *kmem_name;
28 /* Set if any memory will remain added when the driver will be unloaded. */
29 static bool any_hotremove_failed;
30
dax_kmem_range(struct dev_dax * dev_dax,int i,struct range * r)31 static 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
47 struct dax_kmem_data {
48 const char *res_name;
49 int mgid;
50 struct resource *res[];
51 };
52
53 static DEFINE_MUTEX(kmem_memory_type_lock);
54 static LIST_HEAD(kmem_memory_types);
55
kmem_find_alloc_memory_type(int adist)56 static 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
kmem_put_memory_types(void)62 static void kmem_put_memory_types(void)
63 {
64 guard(mutex)(&kmem_memory_type_lock);
65 mt_put_memory_types(&kmem_memory_types);
66 }
67
dev_dax_kmem_probe(struct dev_dax * dev_dax)68 static int dev_dax_kmem_probe(struct dev_dax *dev_dax)
69 {
70 struct device *dev = &dev_dax->dev;
71 unsigned long total_len = 0, orig_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 orig_len += range_len(&dev_dax->ranges[i].range);
101 rc = dax_kmem_range(dev_dax, i, &range);
102 if (rc) {
103 dev_info(dev, "mapping%d: %#llx-%#llx too small after alignment\n",
104 i, range.start, range.end);
105 continue;
106 }
107 total_len += range_len(&range);
108 }
109
110 if (!total_len) {
111 dev_warn(dev, "rejecting DAX region without any memory after alignment\n");
112 return -EINVAL;
113 } else if (total_len != orig_len) {
114 char buf[16];
115
116 string_get_size(orig_len - total_len, 1, STRING_UNITS_2,
117 buf, sizeof(buf));
118 dev_warn(dev, "DAX region truncated by %s due to alignment\n", buf);
119 }
120
121 init_node_memory_type(numa_node, mtype);
122
123 rc = -ENOMEM;
124 data = kzalloc(struct_size(data, res, dev_dax->nr_range), GFP_KERNEL);
125 if (!data)
126 goto err_dax_kmem_data;
127
128 data->res_name = kstrdup(dev_name(dev), GFP_KERNEL);
129 if (!data->res_name)
130 goto err_res_name;
131
132 rc = memory_group_register_static(numa_node, PFN_UP(total_len));
133 if (rc < 0)
134 goto err_reg_mgid;
135 data->mgid = rc;
136
137 for (i = 0; i < dev_dax->nr_range; i++) {
138 struct resource *res;
139 struct range range;
140
141 rc = dax_kmem_range(dev_dax, i, &range);
142 if (rc)
143 continue;
144
145 /* Region is permanently reserved if hotremove fails. */
146 res = request_mem_region(range.start, range_len(&range), data->res_name);
147 if (!res) {
148 dev_warn(dev, "mapping%d: %#llx-%#llx could not reserve region\n",
149 i, range.start, range.end);
150 /*
151 * Once some memory has been onlined we can't
152 * assume that it can be un-onlined safely.
153 */
154 if (mapped)
155 continue;
156 rc = -EBUSY;
157 goto err_request_mem;
158 }
159 data->res[i] = res;
160
161 /*
162 * Set flags appropriate for System RAM. Leave ..._BUSY clear
163 * so that add_memory() can add a child resource. Do not
164 * inherit flags from the parent since it may set new flags
165 * unknown to us that will break add_memory() below.
166 */
167 res->flags = IORESOURCE_SYSTEM_RAM;
168
169 mhp_flags = MHP_NID_IS_MGID;
170 if (dev_dax->memmap_on_memory)
171 mhp_flags |= MHP_MEMMAP_ON_MEMORY;
172
173 /*
174 * Ensure that future kexec'd kernels will not treat
175 * this as RAM automatically.
176 */
177 rc = add_memory_driver_managed(data->mgid, range.start,
178 range_len(&range), kmem_name, mhp_flags);
179
180 if (rc) {
181 dev_warn(dev, "mapping%d: %#llx-%#llx memory add failed\n",
182 i, range.start, range.end);
183 remove_resource(res);
184 kfree(res);
185 data->res[i] = NULL;
186 if (mapped)
187 continue;
188 goto err_request_mem;
189 }
190 mapped++;
191 }
192
193 dev_set_drvdata(dev, data);
194
195 return 0;
196
197 err_request_mem:
198 memory_group_unregister(data->mgid);
199 err_reg_mgid:
200 kfree(data->res_name);
201 err_res_name:
202 kfree(data);
203 err_dax_kmem_data:
204 clear_node_memory_type(numa_node, mtype);
205 return rc;
206 }
207
208 #ifdef CONFIG_MEMORY_HOTREMOVE
dev_dax_kmem_remove(struct dev_dax * dev_dax)209 static void dev_dax_kmem_remove(struct dev_dax *dev_dax)
210 {
211 int i, success = 0;
212 int node = dev_dax->target_node;
213 struct device *dev = &dev_dax->dev;
214 struct dax_kmem_data *data = dev_get_drvdata(dev);
215
216 /*
217 * We have one shot for removing memory, if some memory blocks were not
218 * offline prior to calling this function remove_memory() will fail, and
219 * there is no way to hotremove this memory until reboot because device
220 * unbind will succeed even if we return failure.
221 */
222 for (i = 0; i < dev_dax->nr_range; i++) {
223 struct range range;
224 int rc;
225
226 rc = dax_kmem_range(dev_dax, i, &range);
227 if (rc)
228 continue;
229
230 rc = remove_memory(range.start, range_len(&range));
231 if (rc == 0) {
232 remove_resource(data->res[i]);
233 kfree(data->res[i]);
234 data->res[i] = NULL;
235 success++;
236 continue;
237 }
238 any_hotremove_failed = true;
239 dev_err(dev,
240 "mapping%d: %#llx-%#llx cannot be hotremoved until the next reboot\n",
241 i, range.start, range.end);
242 }
243
244 if (success >= dev_dax->nr_range) {
245 memory_group_unregister(data->mgid);
246 kfree(data->res_name);
247 kfree(data);
248 dev_set_drvdata(dev, NULL);
249 /*
250 * Clear the memtype association on successful unplug.
251 * If not, we have memory blocks left which can be
252 * offlined/onlined later. We need to keep memory_dev_type
253 * for that. This implies this reference will be around
254 * till next reboot.
255 */
256 clear_node_memory_type(node, NULL);
257 }
258 }
259 #else
dev_dax_kmem_remove(struct dev_dax * dev_dax)260 static void dev_dax_kmem_remove(struct dev_dax *dev_dax)
261 {
262 /*
263 * Without hotremove purposely leak the request_mem_region() for the
264 * device-dax range and return '0' to ->remove() attempts. The removal
265 * of the device from the driver always succeeds, but the region is
266 * permanently pinned as reserved by the unreleased
267 * request_mem_region().
268 */
269 any_hotremove_failed = true;
270 }
271 #endif /* CONFIG_MEMORY_HOTREMOVE */
272
273 static struct dax_device_driver device_dax_kmem_driver = {
274 .probe = dev_dax_kmem_probe,
275 .remove = dev_dax_kmem_remove,
276 .type = DAXDRV_KMEM_TYPE,
277 };
278
dax_kmem_init(void)279 static int __init dax_kmem_init(void)
280 {
281 int rc;
282
283 /* Resource name is permanently allocated if any hotremove fails. */
284 kmem_name = kstrdup_const("System RAM (kmem)", GFP_KERNEL);
285 if (!kmem_name)
286 return -ENOMEM;
287
288 rc = dax_driver_register(&device_dax_kmem_driver);
289 if (rc)
290 goto error_dax_driver;
291
292 return rc;
293
294 error_dax_driver:
295 kmem_put_memory_types();
296 kfree_const(kmem_name);
297 return rc;
298 }
299
dax_kmem_exit(void)300 static void __exit dax_kmem_exit(void)
301 {
302 dax_driver_unregister(&device_dax_kmem_driver);
303 if (!any_hotremove_failed)
304 kfree_const(kmem_name);
305 kmem_put_memory_types();
306 }
307
308 MODULE_AUTHOR("Intel Corporation");
309 MODULE_DESCRIPTION("KMEM DAX: map dax-devices as System-RAM");
310 MODULE_LICENSE("GPL v2");
311 module_init(dax_kmem_init);
312 module_exit(dax_kmem_exit);
313 MODULE_ALIAS_DAX_DEVICE(0);
314