drivers/dax/kmem.c

   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(). */
  18 static const char *kmem_name;
  19 /* Set if any memory will remain added when the driver will be unloaded. */
  20 static bool any_hotremove_failed;
  21
  22 static 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
  38 struct dax_kmem_data {
  39         const char *res_name;
  40         struct resource *res[];
  41 };
  42
  43 static 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
 131 err_request_mem:
 132         kfree(data->res_name);
 133 err_res_name:
 134         kfree(data);
 135         return rc;
 136 }
 137
 138 #ifdef CONFIG_MEMORY_HOTREMOVE
 139 static int 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         return 0;
 181 }
 182 #else
 183 static int dev_dax_kmem_remove(struct dev_dax *dev_dax)
 184 {
 185         /*
 186          * Without hotremove purposely leak the request_mem_region() for the
 187          * device-dax range and return '0' to ->remove() attempts. The removal
 188          * of the device from the driver always succeeds, but the region is
 189          * permanently pinned as reserved by the unreleased
 190          * request_mem_region().
 191          */
 192         any_hotremove_failed = true;
 193         return 0;
 194 }
 195 #endif /* CONFIG_MEMORY_HOTREMOVE */
 196
 197 static struct dax_device_driver device_dax_kmem_driver = {
 198         .probe = dev_dax_kmem_probe,
 199         .remove = dev_dax_kmem_remove,
 200 };
 201
 202 static int __init dax_kmem_init(void)
 203 {
 204         int rc;
 205
 206         /* Resource name is permanently allocated if any hotremove fails. */
 207         kmem_name = kstrdup_const("System RAM (kmem)", GFP_KERNEL);
 208         if (!kmem_name)
 209                 return -ENOMEM;
 210
 211         rc = dax_driver_register(&device_dax_kmem_driver);
 212         if (rc)
 213                 kfree_const(kmem_name);
 214         return rc;
 215 }
 216
 217 static void __exit dax_kmem_exit(void)
 218 {
 219         dax_driver_unregister(&device_dax_kmem_driver);
 220         if (!any_hotremove_failed)
 221                 kfree_const(kmem_name);
 222 }
 223
 224 MODULE_AUTHOR("Intel Corporation");
 225 MODULE_LICENSE("GPL v2");
 226 module_init(dax_kmem_init);
 227 module_exit(dax_kmem_exit);
 228 MODULE_ALIAS_DAX_DEVICE(0);