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 "dax-private.h"
15#include "bus.h"
16
17/* Memory resource name used for add_memory_driver_managed(). */
18static const char *kmem_name;
19/* Set if any memory will remain added when the driver will be unloaded. */
20static bool any_hotremove_failed;
21
22int dev_dax_kmem_probe(struct device *dev)
23{
24 struct dev_dax *dev_dax = to_dev_dax(dev);
25 struct resource *res = &dev_dax->region->res;
26 resource_size_t kmem_start;
27 resource_size_t kmem_size;
28 resource_size_t kmem_end;
29 struct resource *new_res;
30 const char *new_res_name;
31 int numa_node;
32 int rc;
33
34 /*
35 * Ensure good NUMA information for the persistent memory.
36 * Without this check, there is a risk that slow memory
37 * could be mixed in a node with faster memory, causing
38 * unavoidable performance issues.
39 */
40 numa_node = dev_dax->target_node;
41 if (numa_node < 0) {
42 dev_warn(dev, "rejecting DAX region %pR with invalid node: %d\n",
43 res, numa_node);
44 return -EINVAL;
45 }
46
47 /* Hotplug starting at the beginning of the next block: */
48 kmem_start = ALIGN(res->start, memory_block_size_bytes());
49
50 kmem_size = resource_size(res);
51 /* Adjust the size down to compensate for moving up kmem_start: */
52 kmem_size -= kmem_start - res->start;
53 /* Align the size down to cover only complete blocks: */
54 kmem_size &= ~(memory_block_size_bytes() - 1);
55 kmem_end = kmem_start + kmem_size;
56
57 new_res_name = kstrdup(dev_name(dev), GFP_KERNEL);
58 if (!new_res_name)
59 return -ENOMEM;
60
61 /* Region is permanently reserved if hotremove fails. */
62 new_res = request_mem_region(kmem_start, kmem_size, new_res_name);
63 if (!new_res) {
64 dev_warn(dev, "could not reserve region [%pa-%pa]\n",
65 &kmem_start, &kmem_end);
66 kfree(new_res_name);
67 return -EBUSY;
68 }
69
70 /*
71 * Set flags appropriate for System RAM. Leave ..._BUSY clear
72 * so that add_memory() can add a child resource. Do not
73 * inherit flags from the parent since it may set new flags
74 * unknown to us that will break add_memory() below.
75 */
76 new_res->flags = IORESOURCE_SYSTEM_RAM;
77
78 /*
79 * Ensure that future kexec'd kernels will not treat this as RAM
80 * automatically.
81 */
82 rc = add_memory_driver_managed(numa_node, new_res->start,
83 resource_size(new_res), kmem_name);
84 if (rc) {
85 release_resource(new_res);
86 kfree(new_res);
87 kfree(new_res_name);
88 return rc;
89 }
90 dev_dax->dax_kmem_res = new_res;
91
92 return 0;
93}
94
95#ifdef CONFIG_MEMORY_HOTREMOVE
96static int dev_dax_kmem_remove(struct device *dev)
97{
98 struct dev_dax *dev_dax = to_dev_dax(dev);
99 struct resource *res = dev_dax->dax_kmem_res;
100 resource_size_t kmem_start = res->start;
101 resource_size_t kmem_size = resource_size(res);
102 const char *res_name = res->name;
103 int rc;
104
105 /*
106 * We have one shot for removing memory, if some memory blocks were not
107 * offline prior to calling this function remove_memory() will fail, and
108 * there is no way to hotremove this memory until reboot because device
109 * unbind will succeed even if we return failure.
110 */
111 rc = remove_memory(dev_dax->target_node, kmem_start, kmem_size);
112 if (rc) {
113 any_hotremove_failed = true;
114 dev_err(dev,
115 "DAX region %pR cannot be hotremoved until the next reboot\n",
116 res);
117 return rc;
118 }
119
120 /* Release and free dax resources */
121 release_resource(res);
122 kfree(res);
123 kfree(res_name);
124 dev_dax->dax_kmem_res = NULL;
125
126 return 0;
127}
128#else
129static int dev_dax_kmem_remove(struct device *dev)
130{
131 /*
132 * Without hotremove purposely leak the request_mem_region() for the
133 * device-dax range and return '0' to ->remove() attempts. The removal
134 * of the device from the driver always succeeds, but the region is
135 * permanently pinned as reserved by the unreleased
136 * request_mem_region().
137 */
138 any_hotremove_failed = true;
139 return 0;
140}
141#endif /* CONFIG_MEMORY_HOTREMOVE */
142
143static struct dax_device_driver device_dax_kmem_driver = {
144 .drv = {
145 .probe = dev_dax_kmem_probe,
146 .remove = dev_dax_kmem_remove,
147 },
148};
149
150static int __init dax_kmem_init(void)
151{
152 int rc;
153
154 /* Resource name is permanently allocated if any hotremove fails. */
155 kmem_name = kstrdup_const("System RAM (kmem)", GFP_KERNEL);
156 if (!kmem_name)
157 return -ENOMEM;
158
159 rc = dax_driver_register(&device_dax_kmem_driver);
160 if (rc)
161 kfree_const(kmem_name);
162 return rc;
163}
164
165static void __exit dax_kmem_exit(void)
166{
167 dax_driver_unregister(&device_dax_kmem_driver);
168 if (!any_hotremove_failed)
169 kfree_const(kmem_name);
170}
171
172MODULE_AUTHOR("Intel Corporation");
173MODULE_LICENSE("GPL v2");
174module_init(dax_kmem_init);
175module_exit(dax_kmem_exit);
176MODULE_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 "dax-private.h"
15#include "bus.h"
16
17/* Memory resource name used for add_memory_driver_managed(). */
18static const char *kmem_name;
19/* Set if any memory will remain added when the driver will be unloaded. */
20static bool any_hotremove_failed;
21
22static int dax_kmem_range(struct dev_dax *dev_dax, int i, struct range *r)
23{
24 struct dev_dax_range *dax_range = &dev_dax->ranges[i];
25 struct range *range = &dax_range->range;
26
27 /* memory-block align the hotplug range */
28 r->start = ALIGN(range->start, memory_block_size_bytes());
29 r->end = ALIGN_DOWN(range->end + 1, memory_block_size_bytes()) - 1;
30 if (r->start >= r->end) {
31 r->start = range->start;
32 r->end = range->end;
33 return -ENOSPC;
34 }
35 return 0;
36}
37
38struct dax_kmem_data {
39 const char *res_name;
40 struct resource *res[];
41};
42
43static int dev_dax_kmem_probe(struct dev_dax *dev_dax)
44{
45 struct device *dev = &dev_dax->dev;
46 struct dax_kmem_data *data;
47 int rc = -ENOMEM;
48 int i, mapped = 0;
49 int numa_node;
50
51 /*
52 * Ensure good NUMA information for the persistent memory.
53 * Without this check, there is a risk that slow memory
54 * could be mixed in a node with faster memory, causing
55 * unavoidable performance issues.
56 */
57 numa_node = dev_dax->target_node;
58 if (numa_node < 0) {
59 dev_warn(dev, "rejecting DAX region with invalid node: %d\n",
60 numa_node);
61 return -EINVAL;
62 }
63
64 data = kzalloc(struct_size(data, res, dev_dax->nr_range), GFP_KERNEL);
65 if (!data)
66 return -ENOMEM;
67
68 data->res_name = kstrdup(dev_name(dev), GFP_KERNEL);
69 if (!data->res_name)
70 goto err_res_name;
71
72 for (i = 0; i < dev_dax->nr_range; i++) {
73 struct resource *res;
74 struct range range;
75
76 rc = dax_kmem_range(dev_dax, i, &range);
77 if (rc) {
78 dev_info(dev, "mapping%d: %#llx-%#llx too small after alignment\n",
79 i, range.start, range.end);
80 continue;
81 }
82
83 /* Region is permanently reserved if hotremove fails. */
84 res = request_mem_region(range.start, range_len(&range), data->res_name);
85 if (!res) {
86 dev_warn(dev, "mapping%d: %#llx-%#llx could not reserve region\n",
87 i, range.start, range.end);
88 /*
89 * Once some memory has been onlined we can't
90 * assume that it can be un-onlined safely.
91 */
92 if (mapped)
93 continue;
94 rc = -EBUSY;
95 goto err_request_mem;
96 }
97 data->res[i] = res;
98
99 /*
100 * Set flags appropriate for System RAM. Leave ..._BUSY clear
101 * so that add_memory() can add a child resource. Do not
102 * inherit flags from the parent since it may set new flags
103 * unknown to us that will break add_memory() below.
104 */
105 res->flags = IORESOURCE_SYSTEM_RAM;
106
107 /*
108 * Ensure that future kexec'd kernels will not treat
109 * this as RAM automatically.
110 */
111 rc = add_memory_driver_managed(numa_node, range.start,
112 range_len(&range), kmem_name, MHP_NONE);
113
114 if (rc) {
115 dev_warn(dev, "mapping%d: %#llx-%#llx memory add failed\n",
116 i, range.start, range.end);
117 release_resource(res);
118 kfree(res);
119 data->res[i] = NULL;
120 if (mapped)
121 continue;
122 goto err_request_mem;
123 }
124 mapped++;
125 }
126
127 dev_set_drvdata(dev, data);
128
129 return 0;
130
131err_request_mem:
132 kfree(data->res_name);
133err_res_name:
134 kfree(data);
135 return rc;
136}
137
138#ifdef CONFIG_MEMORY_HOTREMOVE
139static void dev_dax_kmem_remove(struct dev_dax *dev_dax)
140{
141 int i, success = 0;
142 struct device *dev = &dev_dax->dev;
143 struct dax_kmem_data *data = dev_get_drvdata(dev);
144
145 /*
146 * We have one shot for removing memory, if some memory blocks were not
147 * offline prior to calling this function remove_memory() will fail, and
148 * there is no way to hotremove this memory until reboot because device
149 * unbind will succeed even if we return failure.
150 */
151 for (i = 0; i < dev_dax->nr_range; i++) {
152 struct range range;
153 int rc;
154
155 rc = dax_kmem_range(dev_dax, i, &range);
156 if (rc)
157 continue;
158
159 rc = remove_memory(dev_dax->target_node, range.start,
160 range_len(&range));
161 if (rc == 0) {
162 release_resource(data->res[i]);
163 kfree(data->res[i]);
164 data->res[i] = NULL;
165 success++;
166 continue;
167 }
168 any_hotremove_failed = true;
169 dev_err(dev,
170 "mapping%d: %#llx-%#llx cannot be hotremoved until the next reboot\n",
171 i, range.start, range.end);
172 }
173
174 if (success >= dev_dax->nr_range) {
175 kfree(data->res_name);
176 kfree(data);
177 dev_set_drvdata(dev, NULL);
178 }
179}
180#else
181static void dev_dax_kmem_remove(struct dev_dax *dev_dax)
182{
183 /*
184 * Without hotremove purposely leak the request_mem_region() for the
185 * device-dax range and return '0' to ->remove() attempts. The removal
186 * of the device from the driver always succeeds, but the region is
187 * permanently pinned as reserved by the unreleased
188 * request_mem_region().
189 */
190 any_hotremove_failed = true;
191}
192#endif /* CONFIG_MEMORY_HOTREMOVE */
193
194static struct dax_device_driver device_dax_kmem_driver = {
195 .probe = dev_dax_kmem_probe,
196 .remove = dev_dax_kmem_remove,
197};
198
199static int __init dax_kmem_init(void)
200{
201 int rc;
202
203 /* Resource name is permanently allocated if any hotremove fails. */
204 kmem_name = kstrdup_const("System RAM (kmem)", GFP_KERNEL);
205 if (!kmem_name)
206 return -ENOMEM;
207
208 rc = dax_driver_register(&device_dax_kmem_driver);
209 if (rc)
210 kfree_const(kmem_name);
211 return rc;
212}
213
214static void __exit dax_kmem_exit(void)
215{
216 dax_driver_unregister(&device_dax_kmem_driver);
217 if (!any_hotremove_failed)
218 kfree_const(kmem_name);
219}
220
221MODULE_AUTHOR("Intel Corporation");
222MODULE_LICENSE("GPL v2");
223module_init(dax_kmem_init);
224module_exit(dax_kmem_exit);
225MODULE_ALIAS_DAX_DEVICE(0);