arch/powerpc/platforms/powernv/memtrace.c

   1 /*
   2  * Copyright (C) IBM Corporation, 2014, 2017
   3  * Anton Blanchard, Rashmica Gupta.
   4  *
   5  * This program is free software; you can redistribute it and/or modify
   6  * it under the terms of the GNU General Public License as published by
   7  * the Free Software Foundation; either version 2 of the License, or
   8  * (at your option) any later version.
   9  */
  10
  11 #define pr_fmt(fmt) "memtrace: " fmt
  12
  13 #include <linux/bitops.h>
  14 #include <linux/string.h>
  15 #include <linux/memblock.h>
  16 #include <linux/init.h>
  17 #include <linux/moduleparam.h>
  18 #include <linux/fs.h>
  19 #include <linux/debugfs.h>
  20 #include <linux/slab.h>
  21 #include <linux/memory.h>
  22 #include <linux/memory_hotplug.h>
  23 #include <asm/machdep.h>
  24 #include <asm/debugfs.h>
  25
  26 /* This enables us to keep track of the memory removed from each node. */
  27 struct memtrace_entry {
  28         void *mem;
  29         u64 start;
  30         u64 size;
  31         u32 nid;
  32         struct dentry *dir;
  33         char name[16];
  34 };
  35
  36 static DEFINE_MUTEX(memtrace_mutex);
  37 static u64 memtrace_size;
  38
  39 static struct memtrace_entry *memtrace_array;
  40 static unsigned int memtrace_array_nr;
  41
  42
  43 static ssize_t memtrace_read(struct file *filp, char __user *ubuf,
  44                              size_t count, loff_t *ppos)
  45 {
  46         struct memtrace_entry *ent = filp->private_data;
  47
  48         return simple_read_from_buffer(ubuf, count, ppos, ent->mem, ent->size);
  49 }
  50
  51 static const struct file_operations memtrace_fops = {
  52         .llseek = default_llseek,
  53         .read   = memtrace_read,
  54         .open   = simple_open,
  55 };
  56
  57 static int check_memblock_online(struct memory_block *mem, void *arg)
  58 {
  59         if (mem->state != MEM_ONLINE)
  60                 return -1;
  61
  62         return 0;
  63 }
  64
  65 static int change_memblock_state(struct memory_block *mem, void *arg)
  66 {
  67         unsigned long state = (unsigned long)arg;
  68
  69         mem->state = state;
  70
  71         return 0;
  72 }
  73
  74 static void memtrace_clear_range(unsigned long start_pfn,
  75                                  unsigned long nr_pages)
  76 {
  77         unsigned long pfn;
  78
  79         /*
  80          * As pages are offline, we cannot trust the memmap anymore. As HIGHMEM
  81          * does not apply, avoid passing around "struct page" and use
  82          * clear_page() instead directly.
  83          */
  84         for (pfn = start_pfn; pfn < start_pfn + nr_pages; pfn++) {
  85                 if (IS_ALIGNED(pfn, PAGES_PER_SECTION))
  86                         cond_resched();
  87                 clear_page(__va(PFN_PHYS(pfn)));
  88         }
  89 }
  90
  91 /* called with device_hotplug_lock held */
  92 static bool memtrace_offline_pages(u32 nid, u64 start_pfn, u64 nr_pages)
  93 {
  94         u64 end_pfn = start_pfn + nr_pages - 1;
  95
  96         if (walk_memory_range(start_pfn, end_pfn, NULL,
  97             check_memblock_online))
  98                 return false;
  99
 100         walk_memory_range(start_pfn, end_pfn, (void *)MEM_GOING_OFFLINE,
 101                           change_memblock_state);
 102
 103         if (offline_pages(start_pfn, nr_pages)) {
 104                 walk_memory_range(start_pfn, end_pfn, (void *)MEM_ONLINE,
 105                                   change_memblock_state);
 106                 return false;
 107         }
 108
 109         walk_memory_range(start_pfn, end_pfn, (void *)MEM_OFFLINE,
 110                           change_memblock_state);
 111
 112
 113         return true;
 114 }
 115
 116 static u64 memtrace_alloc_node(u32 nid, u64 size)
 117 {
 118         u64 start_pfn, end_pfn, nr_pages, pfn;
 119         u64 base_pfn;
 120         u64 bytes = memory_block_size_bytes();
 121
 122         if (!node_spanned_pages(nid))
 123                 return 0;
 124
 125         start_pfn = node_start_pfn(nid);
 126         end_pfn = node_end_pfn(nid);
 127         nr_pages = size >> PAGE_SHIFT;
 128
 129         /* Trace memory needs to be aligned to the size */
 130         end_pfn = round_down(end_pfn - nr_pages, nr_pages);
 131
 132         lock_device_hotplug();
 133         for (base_pfn = end_pfn; base_pfn > start_pfn; base_pfn -= nr_pages) {
 134                 if (memtrace_offline_pages(nid, base_pfn, nr_pages) == true) {
 135                         /*
 136                          * Clear the range while we still have a linear
 137                          * mapping.
 138                          */
 139                         memtrace_clear_range(base_pfn, nr_pages);
 140                         /*
 141                          * Remove memory in memory block size chunks so that
 142                          * iomem resources are always split to the same size and
 143                          * we never try to remove memory that spans two iomem
 144                          * resources.
 145                          */
 146                         end_pfn = base_pfn + nr_pages;
 147                         for (pfn = base_pfn; pfn < end_pfn; pfn += bytes>> PAGE_SHIFT) {
 148                                 __remove_memory(nid, pfn << PAGE_SHIFT, bytes);
 149                         }
 150                         unlock_device_hotplug();
 151                         return base_pfn << PAGE_SHIFT;
 152                 }
 153         }
 154         unlock_device_hotplug();
 155
 156         return 0;
 157 }
 158
 159 static int memtrace_init_regions_runtime(u64 size)
 160 {
 161         u32 nid;
 162         u64 m;
 163
 164         memtrace_array = kcalloc(num_online_nodes(),
 165                                 sizeof(struct memtrace_entry), GFP_KERNEL);
 166         if (!memtrace_array) {
 167                 pr_err("Failed to allocate memtrace_array\n");
 168                 return -EINVAL;
 169         }
 170
 171         for_each_online_node(nid) {
 172                 m = memtrace_alloc_node(nid, size);
 173
 174                 /*
 175                  * A node might not have any local memory, so warn but
 176                  * continue on.
 177                  */
 178                 if (!m) {
 179                         pr_err("Failed to allocate trace memory on node %d\n", nid);
 180                         continue;
 181                 }
 182
 183                 pr_info("Allocated trace memory on node %d at 0x%016llx\n", nid, m);
 184
 185                 memtrace_array[memtrace_array_nr].start = m;
 186                 memtrace_array[memtrace_array_nr].size = size;
 187                 memtrace_array[memtrace_array_nr].nid = nid;
 188                 memtrace_array_nr++;
 189         }
 190
 191         return 0;
 192 }
 193
 194 static struct dentry *memtrace_debugfs_dir;
 195
 196 static int memtrace_init_debugfs(void)
 197 {
 198         int ret = 0;
 199         int i;
 200
 201         for (i = 0; i < memtrace_array_nr; i++) {
 202                 struct dentry *dir;
 203                 struct memtrace_entry *ent = &memtrace_array[i];
 204
 205                 ent->mem = ioremap(ent->start, ent->size);
 206                 /* Warn but continue on */
 207                 if (!ent->mem) {
 208                         pr_err("Failed to map trace memory at 0x%llx\n",
 209                                  ent->start);
 210                         ret = -1;
 211                         continue;
 212                 }
 213
 214                 snprintf(ent->name, 16, "%08x", ent->nid);
 215                 dir = debugfs_create_dir(ent->name, memtrace_debugfs_dir);
 216                 if (!dir) {
 217                         pr_err("Failed to create debugfs directory for node %d\n",
 218                                 ent->nid);
 219                         return -1;
 220                 }
 221
 222                 ent->dir = dir;
 223                 debugfs_create_file("trace", 0400, dir, ent, &memtrace_fops);
 224                 debugfs_create_x64("start", 0400, dir, &ent->start);
 225                 debugfs_create_x64("size", 0400, dir, &ent->size);
 226         }
 227
 228         return ret;
 229 }
 230
 231 static int online_mem_block(struct memory_block *mem, void *arg)
 232 {
 233         return device_online(&mem->dev);
 234 }
 235
 236 /*
 237  * Iterate through the chunks of memory we have removed from the kernel
 238  * and attempt to add them back to the kernel.
 239  */
 240 static int memtrace_online(void)
 241 {
 242         int i, ret = 0;
 243         struct memtrace_entry *ent;
 244
 245         for (i = memtrace_array_nr - 1; i >= 0; i--) {
 246                 ent = &memtrace_array[i];
 247
 248                 /* We have onlined this chunk previously */
 249                 if (ent->nid == -1)
 250                         continue;
 251
 252                 /* Remove from io mappings */
 253                 if (ent->mem) {
 254                         iounmap(ent->mem);
 255                         ent->mem = 0;
 256                 }
 257
 258                 if (add_memory(ent->nid, ent->start, ent->size)) {
 259                         pr_err("Failed to add trace memory to node %d\n",
 260                                 ent->nid);
 261                         ret += 1;
 262                         continue;
 263                 }
 264
 265                 /*
 266                  * If kernel isn't compiled with the auto online option
 267                  * we need to online the memory ourselves.
 268                  */
 269                 if (!memhp_auto_online) {
 270                         lock_device_hotplug();
 271                         walk_memory_range(PFN_DOWN(ent->start),
 272                                           PFN_UP(ent->start + ent->size - 1),
 273                                           NULL, online_mem_block);
 274                         unlock_device_hotplug();
 275                 }
 276
 277                 /*
 278                  * Memory was added successfully so clean up references to it
 279                  * so on reentry we can tell that this chunk was added.
 280                  */
 281                 debugfs_remove_recursive(ent->dir);
 282                 pr_info("Added trace memory back to node %d\n", ent->nid);
 283                 ent->size = ent->start = ent->nid = -1;
 284         }
 285         if (ret)
 286                 return ret;
 287
 288         /* If all chunks of memory were added successfully, reset globals */
 289         kfree(memtrace_array);
 290         memtrace_array = NULL;
 291         memtrace_size = 0;
 292         memtrace_array_nr = 0;
 293         return 0;
 294 }
 295
 296 static int memtrace_enable_set(void *data, u64 val)
 297 {
 298         int rc = -EAGAIN;
 299         u64 bytes;
 300
 301         /*
 302          * Don't attempt to do anything if size isn't aligned to a memory
 303          * block or equal to zero.
 304          */
 305         bytes = memory_block_size_bytes();
 306         if (val & (bytes - 1)) {
 307                 pr_err("Value must be aligned with 0x%llx\n", bytes);
 308                 return -EINVAL;
 309         }
 310
 311         mutex_lock(&memtrace_mutex);
 312
 313         /* Re-add/online previously removed/offlined memory */
 314         if (memtrace_size) {
 315                 if (memtrace_online())
 316                         goto out_unlock;
 317         }
 318
 319         if (!val) {
 320                 rc = 0;
 321                 goto out_unlock;
 322         }
 323
 324         /* Offline and remove memory */
 325         if (memtrace_init_regions_runtime(val))
 326                 goto out_unlock;
 327
 328         if (memtrace_init_debugfs())
 329                 goto out_unlock;
 330
 331         memtrace_size = val;
 332         rc = 0;
 333 out_unlock:
 334         mutex_unlock(&memtrace_mutex);
 335         return rc;
 336 }
 337
 338 static int memtrace_enable_get(void *data, u64 *val)
 339 {
 340         *val = memtrace_size;
 341         return 0;
 342 }
 343
 344 DEFINE_SIMPLE_ATTRIBUTE(memtrace_init_fops, memtrace_enable_get,
 345                                         memtrace_enable_set, "0x%016llx\n");
 346
 347 static int memtrace_init(void)
 348 {
 349         memtrace_debugfs_dir = debugfs_create_dir("memtrace",
 350                                                   powerpc_debugfs_root);
 351         if (!memtrace_debugfs_dir)
 352                 return -1;
 353
 354         debugfs_create_file("enable", 0600, memtrace_debugfs_dir,
 355                             NULL, &memtrace_init_fops);
 356
 357         return 0;
 358 }
 359 machine_device_initcall(powernv, memtrace_init);