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 <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(). */
  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
  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
  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
  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
  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;
  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
 190 err_request_mem:
 191         memory_group_unregister(data->mgid);
 192 err_reg_mgid:
 193         kfree(data->res_name);
 194 err_res_name:
 195         kfree(data);
 196 err_dax_kmem_data:
 197         clear_node_memory_type(numa_node, mtype);
 198         return rc;
 199 }
 200
 201 #ifdef CONFIG_MEMORY_HOTREMOVE
 202 static 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
 253 static 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
 266 static 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
 272 static 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
 287 error_dax_driver:
 288         kmem_put_memory_types();
 289         kfree_const(kmem_name);
 290         return rc;
 291 }
 292
 293 static 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
 301 MODULE_AUTHOR("Intel Corporation");
 302 MODULE_DESCRIPTION("KMEM DAX: map dax-devices as System-RAM");
 303 MODULE_LICENSE("GPL v2");
 304 module_init(dax_kmem_init);
 305 module_exit(dax_kmem_exit);
 306 MODULE_ALIAS_DAX_DEVICE(0);