arch/arm64/kernel/hibernate.c

   1 // SPDX-License-Identifier: GPL-2.0-only
   2 /*:
   3  * Hibernate support specific for ARM64
   4  *
   5  * Derived from work on ARM hibernation support by:
   6  *
   7  * Ubuntu project, hibernation support for mach-dove
   8  * Copyright (C) 2010 Nokia Corporation (Hiroshi Doyu)
   9  * Copyright (C) 2010 Texas Instruments, Inc. (Teerth Reddy et al.)
  10  * Copyright (C) 2006 Rafael J. Wysocki <rjw@sisk.pl>
  11  */
  12 #define pr_fmt(x) "hibernate: " x
  13 #include <linux/cpu.h>
  14 #include <linux/kvm_host.h>
  15 #include <linux/pm.h>
  16 #include <linux/sched.h>
  17 #include <linux/suspend.h>
  18 #include <linux/utsname.h>
  19
  20 #include <asm/barrier.h>
  21 #include <asm/cacheflush.h>
  22 #include <asm/cputype.h>
  23 #include <asm/daifflags.h>
  24 #include <asm/irqflags.h>
  25 #include <asm/kexec.h>
  26 #include <asm/memory.h>
  27 #include <asm/mmu_context.h>
  28 #include <asm/mte.h>
  29 #include <asm/sections.h>
  30 #include <asm/smp.h>
  31 #include <asm/smp_plat.h>
  32 #include <asm/suspend.h>
  33 #include <asm/sysreg.h>
  34 #include <asm/trans_pgd.h>
  35 #include <asm/virt.h>
  36
  37 /*
  38  * Hibernate core relies on this value being 0 on resume, and marks it
  39  * __nosavedata assuming it will keep the resume kernel's '0' value. This
  40  * doesn't happen with either KASLR.
  41  *
  42  * defined as "__visible int in_suspend __nosavedata" in
  43  * kernel/power/hibernate.c
  44  */
  45 extern int in_suspend;
  46
  47 /* Do we need to reset el2? */
  48 #define el2_reset_needed() (is_hyp_nvhe())
  49
  50 /* hyp-stub vectors, used to restore el2 during resume from hibernate. */
  51 extern char __hyp_stub_vectors[];
  52
  53 /*
  54  * The logical cpu number we should resume on, initialised to a non-cpu
  55  * number.
  56  */
  57 static int sleep_cpu = -EINVAL;
  58
  59 /*
  60  * Values that may not change over hibernate/resume. We put the build number
  61  * and date in here so that we guarantee not to resume with a different
  62  * kernel.
  63  */
  64 struct arch_hibernate_hdr_invariants {
  65         char            uts_version[__NEW_UTS_LEN + 1];
  66 };
  67
  68 /* These values need to be know across a hibernate/restore. */
  69 static struct arch_hibernate_hdr {
  70         struct arch_hibernate_hdr_invariants invariants;
  71
  72         /* These are needed to find the relocated kernel if built with kaslr */
  73         phys_addr_t     ttbr1_el1;
  74         void            (*reenter_kernel)(void);
  75
  76         /*
  77          * We need to know where the __hyp_stub_vectors are after restore to
  78          * re-configure el2.
  79          */
  80         phys_addr_t     __hyp_stub_vectors;
  81
  82         u64             sleep_cpu_mpidr;
  83 } resume_hdr;
  84
  85 static inline void arch_hdr_invariants(struct arch_hibernate_hdr_invariants *i)
  86 {
  87         memset(i, 0, sizeof(*i));
  88         memcpy(i->uts_version, init_utsname()->version, sizeof(i->uts_version));
  89 }
  90
  91 int pfn_is_nosave(unsigned long pfn)
  92 {
  93         unsigned long nosave_begin_pfn = sym_to_pfn(&__nosave_begin);
  94         unsigned long nosave_end_pfn = sym_to_pfn(&__nosave_end - 1);
  95
  96         return ((pfn >= nosave_begin_pfn) && (pfn <= nosave_end_pfn)) ||
  97                 crash_is_nosave(pfn);
  98 }
  99
 100 void notrace save_processor_state(void)
 101 {
 102         WARN_ON(num_online_cpus() != 1);
 103 }
 104
 105 void notrace restore_processor_state(void)
 106 {
 107 }
 108
 109 int arch_hibernation_header_save(void *addr, unsigned int max_size)
 110 {
 111         struct arch_hibernate_hdr *hdr = addr;
 112
 113         if (max_size < sizeof(*hdr))
 114                 return -EOVERFLOW;
 115
 116         arch_hdr_invariants(&hdr->invariants);
 117         hdr->ttbr1_el1          = __pa_symbol(swapper_pg_dir);
 118         hdr->reenter_kernel     = _cpu_resume;
 119
 120         /* We can't use __hyp_get_vectors() because kvm may still be loaded */
 121         if (el2_reset_needed())
 122                 hdr->__hyp_stub_vectors = __pa_symbol(__hyp_stub_vectors);
 123         else
 124                 hdr->__hyp_stub_vectors = 0;
 125
 126         /* Save the mpidr of the cpu we called cpu_suspend() on... */
 127         if (sleep_cpu < 0) {
 128                 pr_err("Failing to hibernate on an unknown CPU.\n");
 129                 return -ENODEV;
 130         }
 131         hdr->sleep_cpu_mpidr = cpu_logical_map(sleep_cpu);
 132         pr_info("Hibernating on CPU %d [mpidr:0x%llx]\n", sleep_cpu,
 133                 hdr->sleep_cpu_mpidr);
 134
 135         return 0;
 136 }
 137 EXPORT_SYMBOL(arch_hibernation_header_save);
 138
 139 int arch_hibernation_header_restore(void *addr)
 140 {
 141         int ret;
 142         struct arch_hibernate_hdr_invariants invariants;
 143         struct arch_hibernate_hdr *hdr = addr;
 144
 145         arch_hdr_invariants(&invariants);
 146         if (memcmp(&hdr->invariants, &invariants, sizeof(invariants))) {
 147                 pr_crit("Hibernate image not generated by this kernel!\n");
 148                 return -EINVAL;
 149         }
 150
 151         sleep_cpu = get_logical_index(hdr->sleep_cpu_mpidr);
 152         pr_info("Hibernated on CPU %d [mpidr:0x%llx]\n", sleep_cpu,
 153                 hdr->sleep_cpu_mpidr);
 154         if (sleep_cpu < 0) {
 155                 pr_crit("Hibernated on a CPU not known to this kernel!\n");
 156                 sleep_cpu = -EINVAL;
 157                 return -EINVAL;
 158         }
 159
 160         ret = bringup_hibernate_cpu(sleep_cpu);
 161         if (ret) {
 162                 sleep_cpu = -EINVAL;
 163                 return ret;
 164         }
 165
 166         resume_hdr = *hdr;
 167
 168         return 0;
 169 }
 170 EXPORT_SYMBOL(arch_hibernation_header_restore);
 171
 172 static void *hibernate_page_alloc(void *arg)
 173 {
 174         return (void *)get_safe_page((__force gfp_t)(unsigned long)arg);
 175 }
 176
 177 /*
 178  * Copies length bytes, starting at src_start into an new page,
 179  * perform cache maintenance, then maps it at the specified address low
 180  * address as executable.
 181  *
 182  * This is used by hibernate to copy the code it needs to execute when
 183  * overwriting the kernel text. This function generates a new set of page
 184  * tables, which it loads into ttbr0.
 185  *
 186  * Length is provided as we probably only want 4K of data, even on a 64K
 187  * page system.
 188  */
 189 static int create_safe_exec_page(void *src_start, size_t length,
 190                                  phys_addr_t *phys_dst_addr)
 191 {
 192         struct trans_pgd_info trans_info = {
 193                 .trans_alloc_page       = hibernate_page_alloc,
 194                 .trans_alloc_arg        = (__force void *)GFP_ATOMIC,
 195         };
 196
 197         void *page = (void *)get_safe_page(GFP_ATOMIC);
 198         phys_addr_t trans_ttbr0;
 199         unsigned long t0sz;
 200         int rc;
 201
 202         if (!page)
 203                 return -ENOMEM;
 204
 205         memcpy(page, src_start, length);
 206         caches_clean_inval_pou((unsigned long)page, (unsigned long)page + length);
 207         rc = trans_pgd_idmap_page(&trans_info, &trans_ttbr0, &t0sz, page);
 208         if (rc)
 209                 return rc;
 210
 211         cpu_install_ttbr0(trans_ttbr0, t0sz);
 212         *phys_dst_addr = virt_to_phys(page);
 213
 214         return 0;
 215 }
 216
 217 #ifdef CONFIG_ARM64_MTE
 218
 219 static DEFINE_XARRAY(mte_pages);
 220
 221 static int save_tags(struct page *page, unsigned long pfn)
 222 {
 223         void *tag_storage, *ret;
 224
 225         tag_storage = mte_allocate_tag_storage();
 226         if (!tag_storage)
 227                 return -ENOMEM;
 228
 229         mte_save_page_tags(page_address(page), tag_storage);
 230
 231         ret = xa_store(&mte_pages, pfn, tag_storage, GFP_KERNEL);
 232         if (WARN(xa_is_err(ret), "Failed to store MTE tags")) {
 233                 mte_free_tag_storage(tag_storage);
 234                 return xa_err(ret);
 235         } else if (WARN(ret, "swsusp: %s: Duplicate entry", __func__)) {
 236                 mte_free_tag_storage(ret);
 237         }
 238
 239         return 0;
 240 }
 241
 242 static void swsusp_mte_free_storage(void)
 243 {
 244         XA_STATE(xa_state, &mte_pages, 0);
 245         void *tags;
 246
 247         xa_lock(&mte_pages);
 248         xas_for_each(&xa_state, tags, ULONG_MAX) {
 249                 mte_free_tag_storage(tags);
 250         }
 251         xa_unlock(&mte_pages);
 252
 253         xa_destroy(&mte_pages);
 254 }
 255
 256 static int swsusp_mte_save_tags(void)
 257 {
 258         struct zone *zone;
 259         unsigned long pfn, max_zone_pfn;
 260         int ret = 0;
 261         int n = 0;
 262
 263         if (!system_supports_mte())
 264                 return 0;
 265
 266         for_each_populated_zone(zone) {
 267                 max_zone_pfn = zone_end_pfn(zone);
 268                 for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++) {
 269                         struct page *page = pfn_to_online_page(pfn);
 270
 271                         if (!page)
 272                                 continue;
 273
 274                         if (!test_bit(PG_mte_tagged, &page->flags))
 275                                 continue;
 276
 277                         ret = save_tags(page, pfn);
 278                         if (ret) {
 279                                 swsusp_mte_free_storage();
 280                                 goto out;
 281                         }
 282
 283                         n++;
 284                 }
 285         }
 286         pr_info("Saved %d MTE pages\n", n);
 287
 288 out:
 289         return ret;
 290 }
 291
 292 static void swsusp_mte_restore_tags(void)
 293 {
 294         XA_STATE(xa_state, &mte_pages, 0);
 295         int n = 0;
 296         void *tags;
 297
 298         xa_lock(&mte_pages);
 299         xas_for_each(&xa_state, tags, ULONG_MAX) {
 300                 unsigned long pfn = xa_state.xa_index;
 301                 struct page *page = pfn_to_online_page(pfn);
 302
 303                 mte_restore_page_tags(page_address(page), tags);
 304
 305                 mte_free_tag_storage(tags);
 306                 n++;
 307         }
 308         xa_unlock(&mte_pages);
 309
 310         pr_info("Restored %d MTE pages\n", n);
 311
 312         xa_destroy(&mte_pages);
 313 }
 314
 315 #else   /* CONFIG_ARM64_MTE */
 316
 317 static int swsusp_mte_save_tags(void)
 318 {
 319         return 0;
 320 }
 321
 322 static void swsusp_mte_restore_tags(void)
 323 {
 324 }
 325
 326 #endif  /* CONFIG_ARM64_MTE */
 327
 328 int swsusp_arch_suspend(void)
 329 {
 330         int ret = 0;
 331         unsigned long flags;
 332         struct sleep_stack_data state;
 333
 334         if (cpus_are_stuck_in_kernel()) {
 335                 pr_err("Can't hibernate: no mechanism to offline secondary CPUs.\n");
 336                 return -EBUSY;
 337         }
 338
 339         flags = local_daif_save();
 340
 341         if (__cpu_suspend_enter(&state)) {
 342                 /* make the crash dump kernel image visible/saveable */
 343                 crash_prepare_suspend();
 344
 345                 ret = swsusp_mte_save_tags();
 346                 if (ret)
 347                         return ret;
 348
 349                 sleep_cpu = smp_processor_id();
 350                 ret = swsusp_save();
 351         } else {
 352                 /* Clean kernel core startup/idle code to PoC*/
 353                 dcache_clean_inval_poc((unsigned long)__mmuoff_data_start,
 354                                     (unsigned long)__mmuoff_data_end);
 355                 dcache_clean_inval_poc((unsigned long)__idmap_text_start,
 356                                     (unsigned long)__idmap_text_end);
 357
 358                 /* Clean kvm setup code to PoC? */
 359                 if (el2_reset_needed()) {
 360                         dcache_clean_inval_poc(
 361                                 (unsigned long)__hyp_idmap_text_start,
 362                                 (unsigned long)__hyp_idmap_text_end);
 363                         dcache_clean_inval_poc((unsigned long)__hyp_text_start,
 364                                             (unsigned long)__hyp_text_end);
 365                 }
 366
 367                 swsusp_mte_restore_tags();
 368
 369                 /* make the crash dump kernel image protected again */
 370                 crash_post_resume();
 371
 372                 /*
 373                  * Tell the hibernation core that we've just restored
 374                  * the memory
 375                  */
 376                 in_suspend = 0;
 377
 378                 sleep_cpu = -EINVAL;
 379                 __cpu_suspend_exit();
 380
 381                 /*
 382                  * Just in case the boot kernel did turn the SSBD
 383                  * mitigation off behind our back, let's set the state
 384                  * to what we expect it to be.
 385                  */
 386                 spectre_v4_enable_mitigation(NULL);
 387         }
 388
 389         local_daif_restore(flags);
 390
 391         return ret;
 392 }
 393
 394 /*
 395  * Setup then Resume from the hibernate image using swsusp_arch_suspend_exit().
 396  *
 397  * Memory allocated by get_safe_page() will be dealt with by the hibernate code,
 398  * we don't need to free it here.
 399  */
 400 int swsusp_arch_resume(void)
 401 {
 402         int rc;
 403         void *zero_page;
 404         size_t exit_size;
 405         pgd_t *tmp_pg_dir;
 406         phys_addr_t el2_vectors;
 407         void __noreturn (*hibernate_exit)(phys_addr_t, phys_addr_t, void *,
 408                                           void *, phys_addr_t, phys_addr_t);
 409         struct trans_pgd_info trans_info = {
 410                 .trans_alloc_page       = hibernate_page_alloc,
 411                 .trans_alloc_arg        = (void *)GFP_ATOMIC,
 412         };
 413
 414         /*
 415          * Restoring the memory image will overwrite the ttbr1 page tables.
 416          * Create a second copy of just the linear map, and use this when
 417          * restoring.
 418          */
 419         rc = trans_pgd_create_copy(&trans_info, &tmp_pg_dir, PAGE_OFFSET,
 420                                    PAGE_END);
 421         if (rc)
 422                 return rc;
 423
 424         /*
 425          * We need a zero page that is zero before & after resume in order
 426          * to break before make on the ttbr1 page tables.
 427          */
 428         zero_page = (void *)get_safe_page(GFP_ATOMIC);
 429         if (!zero_page) {
 430                 pr_err("Failed to allocate zero page.\n");
 431                 return -ENOMEM;
 432         }
 433
 434         if (el2_reset_needed()) {
 435                 rc = trans_pgd_copy_el2_vectors(&trans_info, &el2_vectors);
 436                 if (rc) {
 437                         pr_err("Failed to setup el2 vectors\n");
 438                         return rc;
 439                 }
 440         }
 441
 442         exit_size = __hibernate_exit_text_end - __hibernate_exit_text_start;
 443         /*
 444          * Copy swsusp_arch_suspend_exit() to a safe page. This will generate
 445          * a new set of ttbr0 page tables and load them.
 446          */
 447         rc = create_safe_exec_page(__hibernate_exit_text_start, exit_size,
 448                                    (phys_addr_t *)&hibernate_exit);
 449         if (rc) {
 450                 pr_err("Failed to create safe executable page for hibernate_exit code.\n");
 451                 return rc;
 452         }
 453
 454         /*
 455          * KASLR will cause the el2 vectors to be in a different location in
 456          * the resumed kernel. Load hibernate's temporary copy into el2.
 457          *
 458          * We can skip this step if we booted at EL1, or are running with VHE.
 459          */
 460         if (el2_reset_needed())
 461                 __hyp_set_vectors(el2_vectors);
 462
 463         hibernate_exit(virt_to_phys(tmp_pg_dir), resume_hdr.ttbr1_el1,
 464                        resume_hdr.reenter_kernel, restore_pblist,
 465                        resume_hdr.__hyp_stub_vectors, virt_to_phys(zero_page));
 466
 467         return 0;
 468 }
 469
 470 int hibernate_resume_nonboot_cpu_disable(void)
 471 {
 472         if (sleep_cpu < 0) {
 473                 pr_err("Failing to resume from hibernate on an unknown CPU.\n");
 474                 return -ENODEV;
 475         }
 476
 477         return freeze_secondary_cpus(sleep_cpu);
 478 }