arch/powerpc/platforms/powernv/memtrace.c

   1 // SPDX-License-Identifier: GPL-2.0-or-later
   2 /*
   3  * Copyright (C) IBM Corporation, 2014, 2017
   4  * Anton Blanchard, Rashmica Gupta.
   5  */
   6
   7 #define pr_fmt(fmt) "memtrace: " fmt
   8
   9 #include <linux/bitops.h>
  10 #include <linux/string.h>
  11 #include <linux/memblock.h>
  12 #include <linux/init.h>
  13 #include <linux/moduleparam.h>
  14 #include <linux/fs.h>
  15 #include <linux/debugfs.h>
  16 #include <linux/slab.h>
  17 #include <linux/memory.h>
  18 #include <linux/memory_hotplug.h>
  19 #include <linux/numa.h>
  20 #include <asm/machdep.h>
  21 #include <asm/debugfs.h>
  22
  23 /* This enables us to keep track of the memory removed from each node. */
  24 struct memtrace_entry {
  25         void *mem;
  26         u64 start;
  27         u64 size;
  28         u32 nid;
  29         struct dentry *dir;
  30         char name[16];
  31 };
  32
  33 static DEFINE_MUTEX(memtrace_mutex);
  34 static u64 memtrace_size;
  35
  36 static struct memtrace_entry *memtrace_array;
  37 static unsigned int memtrace_array_nr;
  38
  39
  40 static ssize_t memtrace_read(struct file *filp, char __user *ubuf,
  41                              size_t count, loff_t *ppos)
  42 {
  43         struct memtrace_entry *ent = filp->private_data;
  44
  45         return simple_read_from_buffer(ubuf, count, ppos, ent->mem, ent->size);
  46 }
  47
  48 static const struct file_operations memtrace_fops = {
  49         .llseek = default_llseek,
  50         .read   = memtrace_read,
  51         .open   = simple_open,
  52 };
  53
  54 static int check_memblock_online(struct memory_block *mem, void *arg)
  55 {
  56         if (mem->state != MEM_ONLINE)
  57                 return -1;
  58
  59         return 0;
  60 }
  61
  62 static int change_memblock_state(struct memory_block *mem, void *arg)
  63 {
  64         unsigned long state = (unsigned long)arg;
  65
  66         mem->state = state;
  67
  68         return 0;
  69 }
  70
  71 static void memtrace_clear_range(unsigned long start_pfn,
  72                                  unsigned long nr_pages)
  73 {
  74         unsigned long pfn;
  75
  76         /*
  77          * As pages are offline, we cannot trust the memmap anymore. As HIGHMEM
  78          * does not apply, avoid passing around "struct page" and use
  79          * clear_page() instead directly.
  80          */
  81         for (pfn = start_pfn; pfn < start_pfn + nr_pages; pfn++) {
  82                 if (IS_ALIGNED(pfn, PAGES_PER_SECTION))
  83                         cond_resched();
  84                 clear_page(__va(PFN_PHYS(pfn)));
  85         }
  86 }
  87
  88 /* called with device_hotplug_lock held */
  89 static bool memtrace_offline_pages(u32 nid, u64 start_pfn, u64 nr_pages)
  90 {
  91         const unsigned long start = PFN_PHYS(start_pfn);
  92         const unsigned long size = PFN_PHYS(nr_pages);
  93
  94         if (walk_memory_blocks(start, size, NULL, check_memblock_online))
  95                 return false;
  96
  97         walk_memory_blocks(start, size, (void *)MEM_GOING_OFFLINE,
  98                            change_memblock_state);
  99
 100         if (offline_pages(start_pfn, nr_pages)) {
 101                 walk_memory_blocks(start, size, (void *)MEM_ONLINE,
 102                                    change_memblock_state);
 103                 return false;
 104         }
 105
 106         walk_memory_blocks(start, size, (void *)MEM_OFFLINE,
 107                            change_memblock_state);
 108
 109
 110         return true;
 111 }
 112
 113 static u64 memtrace_alloc_node(u32 nid, u64 size)
 114 {
 115         u64 start_pfn, end_pfn, nr_pages, pfn;
 116         u64 base_pfn;
 117         u64 bytes = memory_block_size_bytes();
 118
 119         if (!node_spanned_pages(nid))
 120                 return 0;
 121
 122         start_pfn = node_start_pfn(nid);
 123         end_pfn = node_end_pfn(nid);
 124         nr_pages = size >> PAGE_SHIFT;
 125
 126         /* Trace memory needs to be aligned to the size */
 127         end_pfn = round_down(end_pfn - nr_pages, nr_pages);
 128
 129         lock_device_hotplug();
 130         for (base_pfn = end_pfn; base_pfn > start_pfn; base_pfn -= nr_pages) {
 131                 if (memtrace_offline_pages(nid, base_pfn, nr_pages) == true) {
 132                         /*
 133                          * Clear the range while we still have a linear
 134                          * mapping.
 135                          */
 136                         memtrace_clear_range(base_pfn, nr_pages);
 137                         /*
 138                          * Remove memory in memory block size chunks so that
 139                          * iomem resources are always split to the same size and
 140                          * we never try to remove memory that spans two iomem
 141                          * resources.
 142                          */
 143                         end_pfn = base_pfn + nr_pages;
 144                         for (pfn = base_pfn; pfn < end_pfn; pfn += bytes>> PAGE_SHIFT) {
 145                                 __remove_memory(nid, pfn << PAGE_SHIFT, bytes);
 146                         }
 147                         unlock_device_hotplug();
 148                         return base_pfn << PAGE_SHIFT;
 149                 }
 150         }
 151         unlock_device_hotplug();
 152
 153         return 0;
 154 }
 155
 156 static int memtrace_init_regions_runtime(u64 size)
 157 {
 158         u32 nid;
 159         u64 m;
 160
 161         memtrace_array = kcalloc(num_online_nodes(),
 162                                 sizeof(struct memtrace_entry), GFP_KERNEL);
 163         if (!memtrace_array) {
 164                 pr_err("Failed to allocate memtrace_array\n");
 165                 return -EINVAL;
 166         }
 167
 168         for_each_online_node(nid) {
 169                 m = memtrace_alloc_node(nid, size);
 170
 171                 /*
 172                  * A node might not have any local memory, so warn but
 173                  * continue on.
 174                  */
 175                 if (!m) {
 176                         pr_err("Failed to allocate trace memory on node %d\n", nid);
 177                         continue;
 178                 }
 179
 180                 pr_info("Allocated trace memory on node %d at 0x%016llx\n", nid, m);
 181
 182                 memtrace_array[memtrace_array_nr].start = m;
 183                 memtrace_array[memtrace_array_nr].size = size;
 184                 memtrace_array[memtrace_array_nr].nid = nid;
 185                 memtrace_array_nr++;
 186         }
 187
 188         return 0;
 189 }
 190
 191 static struct dentry *memtrace_debugfs_dir;
 192
 193 static int memtrace_init_debugfs(void)
 194 {
 195         int ret = 0;
 196         int i;
 197
 198         for (i = 0; i < memtrace_array_nr; i++) {
 199                 struct dentry *dir;
 200                 struct memtrace_entry *ent = &memtrace_array[i];
 201
 202                 ent->mem = ioremap(ent->start, ent->size);
 203                 /* Warn but continue on */
 204                 if (!ent->mem) {
 205                         pr_err("Failed to map trace memory at 0x%llx\n",
 206                                  ent->start);
 207                         ret = -1;
 208                         continue;
 209                 }
 210
 211                 snprintf(ent->name, 16, "%08x", ent->nid);
 212                 dir = debugfs_create_dir(ent->name, memtrace_debugfs_dir);
 213                 if (!dir) {
 214                         pr_err("Failed to create debugfs directory for node %d\n",
 215                                 ent->nid);
 216                         return -1;
 217                 }
 218
 219                 ent->dir = dir;
 220                 debugfs_create_file("trace", 0400, dir, ent, &memtrace_fops);
 221                 debugfs_create_x64("start", 0400, dir, &ent->start);
 222                 debugfs_create_x64("size", 0400, dir, &ent->size);
 223         }
 224
 225         return ret;
 226 }
 227
 228 static int online_mem_block(struct memory_block *mem, void *arg)
 229 {
 230         return device_online(&mem->dev);
 231 }
 232
 233 /*
 234  * Iterate through the chunks of memory we have removed from the kernel
 235  * and attempt to add them back to the kernel.
 236  */
 237 static int memtrace_online(void)
 238 {
 239         int i, ret = 0;
 240         struct memtrace_entry *ent;
 241
 242         for (i = memtrace_array_nr - 1; i >= 0; i--) {
 243                 ent = &memtrace_array[i];
 244
 245                 /* We have onlined this chunk previously */
 246                 if (ent->nid == NUMA_NO_NODE)
 247                         continue;
 248
 249                 /* Remove from io mappings */
 250                 if (ent->mem) {
 251                         iounmap(ent->mem);
 252                         ent->mem = 0;
 253                 }
 254
 255                 if (add_memory(ent->nid, ent->start, ent->size)) {
 256                         pr_err("Failed to add trace memory to node %d\n",
 257                                 ent->nid);
 258                         ret += 1;
 259                         continue;
 260                 }
 261
 262                 /*
 263                  * If kernel isn't compiled with the auto online option
 264                  * we need to online the memory ourselves.
 265                  */
 266                 if (!memhp_auto_online) {
 267                         lock_device_hotplug();
 268                         walk_memory_blocks(ent->start, ent->size, NULL,
 269                                            online_mem_block);
 270                         unlock_device_hotplug();
 271                 }
 272
 273                 /*
 274                  * Memory was added successfully so clean up references to it
 275                  * so on reentry we can tell that this chunk was added.
 276                  */
 277                 debugfs_remove_recursive(ent->dir);
 278                 pr_info("Added trace memory back to node %d\n", ent->nid);
 279                 ent->size = ent->start = ent->nid = NUMA_NO_NODE;
 280         }
 281         if (ret)
 282                 return ret;
 283
 284         /* If all chunks of memory were added successfully, reset globals */
 285         kfree(memtrace_array);
 286         memtrace_array = NULL;
 287         memtrace_size = 0;
 288         memtrace_array_nr = 0;
 289         return 0;
 290 }
 291
 292 static int memtrace_enable_set(void *data, u64 val)
 293 {
 294         int rc = -EAGAIN;
 295         u64 bytes;
 296
 297         /*
 298          * Don't attempt to do anything if size isn't aligned to a memory
 299          * block or equal to zero.
 300          */
 301         bytes = memory_block_size_bytes();
 302         if (val & (bytes - 1)) {
 303                 pr_err("Value must be aligned with 0x%llx\n", bytes);
 304                 return -EINVAL;
 305         }
 306
 307         mutex_lock(&memtrace_mutex);
 308
 309         /* Re-add/online previously removed/offlined memory */
 310         if (memtrace_size) {
 311                 if (memtrace_online())
 312                         goto out_unlock;
 313         }
 314
 315         if (!val) {
 316                 rc = 0;
 317                 goto out_unlock;
 318         }
 319
 320         /* Offline and remove memory */
 321         if (memtrace_init_regions_runtime(val))
 322                 goto out_unlock;
 323
 324         if (memtrace_init_debugfs())
 325                 goto out_unlock;
 326
 327         memtrace_size = val;
 328         rc = 0;
 329 out_unlock:
 330         mutex_unlock(&memtrace_mutex);
 331         return rc;
 332 }
 333
 334 static int memtrace_enable_get(void *data, u64 *val)
 335 {
 336         *val = memtrace_size;
 337         return 0;
 338 }
 339
 340 DEFINE_SIMPLE_ATTRIBUTE(memtrace_init_fops, memtrace_enable_get,
 341                                         memtrace_enable_set, "0x%016llx\n");
 342
 343 static int memtrace_init(void)
 344 {
 345         memtrace_debugfs_dir = debugfs_create_dir("memtrace",
 346                                                   powerpc_debugfs_root);
 347         if (!memtrace_debugfs_dir)
 348                 return -1;
 349
 350         debugfs_create_file("enable", 0600, memtrace_debugfs_dir,
 351                             NULL, &memtrace_init_fops);
 352
 353         return 0;
 354 }
 355 machine_device_initcall(powernv, memtrace_init);