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  *  https://lkml.org/lkml/2010/6/18/4
  11  *  https://lists.linux-foundation.org/pipermail/linux-pm/2010-June/027422.html
  12  *  https://patchwork.kernel.org/patch/96442/
  13  *
  14  * Copyright (C) 2006 Rafael J. Wysocki <rjw@sisk.pl>
  15  */
  16 #define pr_fmt(x) "hibernate: " x
  17 #include <linux/cpu.h>
  18 #include <linux/kvm_host.h>
  19 #include <linux/pm.h>
  20 #include <linux/sched.h>
  21 #include <linux/suspend.h>
  22 #include <linux/utsname.h>
  23
  24 #include <asm/barrier.h>
  25 #include <asm/cacheflush.h>
  26 #include <asm/cputype.h>
  27 #include <asm/daifflags.h>
  28 #include <asm/irqflags.h>
  29 #include <asm/kexec.h>
  30 #include <asm/memory.h>
  31 #include <asm/mmu_context.h>
  32 #include <asm/mte.h>
  33 #include <asm/sections.h>
  34 #include <asm/smp.h>
  35 #include <asm/smp_plat.h>
  36 #include <asm/suspend.h>
  37 #include <asm/sysreg.h>
  38 #include <asm/trans_pgd.h>
  39 #include <asm/virt.h>
  40
  41 /*
  42  * Hibernate core relies on this value being 0 on resume, and marks it
  43  * __nosavedata assuming it will keep the resume kernel's '0' value. This
  44  * doesn't happen with either KASLR.
  45  *
  46  * defined as "__visible int in_suspend __nosavedata" in
  47  * kernel/power/hibernate.c
  48  */
  49 extern int in_suspend;
  50
  51 /* Do we need to reset el2? */
  52 #define el2_reset_needed() (is_hyp_nvhe())
  53
  54 /* hyp-stub vectors, used to restore el2 during resume from hibernate. */
  55 extern char __hyp_stub_vectors[];
  56
  57 /*
  58  * The logical cpu number we should resume on, initialised to a non-cpu
  59  * number.
  60  */
  61 static int sleep_cpu = -EINVAL;
  62
  63 /*
  64  * Values that may not change over hibernate/resume. We put the build number
  65  * and date in here so that we guarantee not to resume with a different
  66  * kernel.
  67  */
  68 struct arch_hibernate_hdr_invariants {
  69         char            uts_version[__NEW_UTS_LEN + 1];
  70 };
  71
  72 /* These values need to be know across a hibernate/restore. */
  73 static struct arch_hibernate_hdr {
  74         struct arch_hibernate_hdr_invariants invariants;
  75
  76         /* These are needed to find the relocated kernel if built with kaslr */
  77         phys_addr_t     ttbr1_el1;
  78         void            (*reenter_kernel)(void);
  79
  80         /*
  81          * We need to know where the __hyp_stub_vectors are after restore to
  82          * re-configure el2.
  83          */
  84         phys_addr_t     __hyp_stub_vectors;
  85
  86         u64             sleep_cpu_mpidr;
  87 } resume_hdr;
  88
  89 static inline void arch_hdr_invariants(struct arch_hibernate_hdr_invariants *i)
  90 {
  91         memset(i, 0, sizeof(*i));
  92         memcpy(i->uts_version, init_utsname()->version, sizeof(i->uts_version));
  93 }
  94
  95 int pfn_is_nosave(unsigned long pfn)
  96 {
  97         unsigned long nosave_begin_pfn = sym_to_pfn(&__nosave_begin);
  98         unsigned long nosave_end_pfn = sym_to_pfn(&__nosave_end - 1);
  99
 100         return ((pfn >= nosave_begin_pfn) && (pfn <= nosave_end_pfn)) ||
 101                 crash_is_nosave(pfn);
 102 }
 103
 104 void notrace save_processor_state(void)
 105 {
 106         WARN_ON(num_online_cpus() != 1);
 107 }
 108
 109 void notrace restore_processor_state(void)
 110 {
 111 }
 112
 113 int arch_hibernation_header_save(void *addr, unsigned int max_size)
 114 {
 115         struct arch_hibernate_hdr *hdr = addr;
 116
 117         if (max_size < sizeof(*hdr))
 118                 return -EOVERFLOW;
 119
 120         arch_hdr_invariants(&hdr->invariants);
 121         hdr->ttbr1_el1          = __pa_symbol(swapper_pg_dir);
 122         hdr->reenter_kernel     = _cpu_resume;
 123
 124         /* We can't use __hyp_get_vectors() because kvm may still be loaded */
 125         if (el2_reset_needed())
 126                 hdr->__hyp_stub_vectors = __pa_symbol(__hyp_stub_vectors);
 127         else
 128                 hdr->__hyp_stub_vectors = 0;
 129
 130         /* Save the mpidr of the cpu we called cpu_suspend() on... */
 131         if (sleep_cpu < 0) {
 132                 pr_err("Failing to hibernate on an unknown CPU.\n");
 133                 return -ENODEV;
 134         }
 135         hdr->sleep_cpu_mpidr = cpu_logical_map(sleep_cpu);
 136         pr_info("Hibernating on CPU %d [mpidr:0x%llx]\n", sleep_cpu,
 137                 hdr->sleep_cpu_mpidr);
 138
 139         return 0;
 140 }
 141 EXPORT_SYMBOL(arch_hibernation_header_save);
 142
 143 int arch_hibernation_header_restore(void *addr)
 144 {
 145         int ret;
 146         struct arch_hibernate_hdr_invariants invariants;
 147         struct arch_hibernate_hdr *hdr = addr;
 148
 149         arch_hdr_invariants(&invariants);
 150         if (memcmp(&hdr->invariants, &invariants, sizeof(invariants))) {
 151                 pr_crit("Hibernate image not generated by this kernel!\n");
 152                 return -EINVAL;
 153         }
 154
 155         sleep_cpu = get_logical_index(hdr->sleep_cpu_mpidr);
 156         pr_info("Hibernated on CPU %d [mpidr:0x%llx]\n", sleep_cpu,
 157                 hdr->sleep_cpu_mpidr);
 158         if (sleep_cpu < 0) {
 159                 pr_crit("Hibernated on a CPU not known to this kernel!\n");
 160                 sleep_cpu = -EINVAL;
 161                 return -EINVAL;
 162         }
 163
 164         ret = bringup_hibernate_cpu(sleep_cpu);
 165         if (ret) {
 166                 sleep_cpu = -EINVAL;
 167                 return ret;
 168         }
 169
 170         resume_hdr = *hdr;
 171
 172         return 0;
 173 }
 174 EXPORT_SYMBOL(arch_hibernation_header_restore);
 175
 176 static void *hibernate_page_alloc(void *arg)
 177 {
 178         return (void *)get_safe_page((__force gfp_t)(unsigned long)arg);
 179 }
 180
 181 /*
 182  * Copies length bytes, starting at src_start into an new page,
 183  * perform cache maintenance, then maps it at the specified address low
 184  * address as executable.
 185  *
 186  * This is used by hibernate to copy the code it needs to execute when
 187  * overwriting the kernel text. This function generates a new set of page
 188  * tables, which it loads into ttbr0.
 189  *
 190  * Length is provided as we probably only want 4K of data, even on a 64K
 191  * page system.
 192  */
 193 static int create_safe_exec_page(void *src_start, size_t length,
 194                                  phys_addr_t *phys_dst_addr)
 195 {
 196         struct trans_pgd_info trans_info = {
 197                 .trans_alloc_page       = hibernate_page_alloc,
 198                 .trans_alloc_arg        = (__force void *)GFP_ATOMIC,
 199         };
 200
 201         void *page = (void *)get_safe_page(GFP_ATOMIC);
 202         phys_addr_t trans_ttbr0;
 203         unsigned long t0sz;
 204         int rc;
 205
 206         if (!page)
 207                 return -ENOMEM;
 208
 209         memcpy(page, src_start, length);
 210         caches_clean_inval_pou((unsigned long)page, (unsigned long)page + length);
 211         rc = trans_pgd_idmap_page(&trans_info, &trans_ttbr0, &t0sz, page);
 212         if (rc)
 213                 return rc;
 214
 215         cpu_install_ttbr0(trans_ttbr0, t0sz);
 216         *phys_dst_addr = virt_to_phys(page);
 217
 218         return 0;
 219 }
 220
 221 #ifdef CONFIG_ARM64_MTE
 222
 223 static DEFINE_XARRAY(mte_pages);
 224
 225 static int save_tags(struct page *page, unsigned long pfn)
 226 {
 227         void *tag_storage, *ret;
 228
 229         tag_storage = mte_allocate_tag_storage();
 230         if (!tag_storage)
 231                 return -ENOMEM;
 232
 233         mte_save_page_tags(page_address(page), tag_storage);
 234
 235         ret = xa_store(&mte_pages, pfn, tag_storage, GFP_KERNEL);
 236         if (WARN(xa_is_err(ret), "Failed to store MTE tags")) {
 237                 mte_free_tag_storage(tag_storage);
 238                 return xa_err(ret);
 239         } else if (WARN(ret, "swsusp: %s: Duplicate entry", __func__)) {
 240                 mte_free_tag_storage(ret);
 241         }
 242
 243         return 0;
 244 }
 245
 246 static void swsusp_mte_free_storage(void)
 247 {
 248         XA_STATE(xa_state, &mte_pages, 0);
 249         void *tags;
 250
 251         xa_lock(&mte_pages);
 252         xas_for_each(&xa_state, tags, ULONG_MAX) {
 253                 mte_free_tag_storage(tags);
 254         }
 255         xa_unlock(&mte_pages);
 256
 257         xa_destroy(&mte_pages);
 258 }
 259
 260 static int swsusp_mte_save_tags(void)
 261 {
 262         struct zone *zone;
 263         unsigned long pfn, max_zone_pfn;
 264         int ret = 0;
 265         int n = 0;
 266
 267         if (!system_supports_mte())
 268                 return 0;
 269
 270         for_each_populated_zone(zone) {
 271                 max_zone_pfn = zone_end_pfn(zone);
 272                 for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++) {
 273                         struct page *page = pfn_to_online_page(pfn);
 274
 275                         if (!page)
 276                                 continue;
 277
 278                         if (!test_bit(PG_mte_tagged, &page->flags))
 279                                 continue;
 280
 281                         ret = save_tags(page, pfn);
 282                         if (ret) {
 283                                 swsusp_mte_free_storage();
 284                                 goto out;
 285                         }
 286
 287                         n++;
 288                 }
 289         }
 290         pr_info("Saved %d MTE pages\n", n);
 291
 292 out:
 293         return ret;
 294 }
 295
 296 static void swsusp_mte_restore_tags(void)
 297 {
 298         XA_STATE(xa_state, &mte_pages, 0);
 299         int n = 0;
 300         void *tags;
 301
 302         xa_lock(&mte_pages);
 303         xas_for_each(&xa_state, tags, ULONG_MAX) {
 304                 unsigned long pfn = xa_state.xa_index;
 305                 struct page *page = pfn_to_online_page(pfn);
 306
 307                 /*
 308                  * It is not required to invoke page_kasan_tag_reset(page)
 309                  * at this point since the tags stored in page->flags are
 310                  * already restored.
 311                  */
 312                 mte_restore_page_tags(page_address(page), tags);
 313
 314                 mte_free_tag_storage(tags);
 315                 n++;
 316         }
 317         xa_unlock(&mte_pages);
 318
 319         pr_info("Restored %d MTE pages\n", n);
 320
 321         xa_destroy(&mte_pages);
 322 }
 323
 324 #else   /* CONFIG_ARM64_MTE */
 325
 326 static int swsusp_mte_save_tags(void)
 327 {
 328         return 0;
 329 }
 330
 331 static void swsusp_mte_restore_tags(void)
 332 {
 333 }
 334
 335 #endif  /* CONFIG_ARM64_MTE */
 336
 337 int swsusp_arch_suspend(void)
 338 {
 339         int ret = 0;
 340         unsigned long flags;
 341         struct sleep_stack_data state;
 342
 343         if (cpus_are_stuck_in_kernel()) {
 344                 pr_err("Can't hibernate: no mechanism to offline secondary CPUs.\n");
 345                 return -EBUSY;
 346         }
 347
 348         flags = local_daif_save();
 349
 350         if (__cpu_suspend_enter(&state)) {
 351                 /* make the crash dump kernel image visible/saveable */
 352                 crash_prepare_suspend();
 353
 354                 ret = swsusp_mte_save_tags();
 355                 if (ret)
 356                         return ret;
 357
 358                 sleep_cpu = smp_processor_id();
 359                 ret = swsusp_save();
 360         } else {
 361                 /* Clean kernel core startup/idle code to PoC*/
 362                 dcache_clean_inval_poc((unsigned long)__mmuoff_data_start,
 363                                     (unsigned long)__mmuoff_data_end);
 364                 dcache_clean_inval_poc((unsigned long)__idmap_text_start,
 365                                     (unsigned long)__idmap_text_end);
 366
 367                 /* Clean kvm setup code to PoC? */
 368                 if (el2_reset_needed()) {
 369                         dcache_clean_inval_poc(
 370                                 (unsigned long)__hyp_idmap_text_start,
 371                                 (unsigned long)__hyp_idmap_text_end);
 372                         dcache_clean_inval_poc((unsigned long)__hyp_text_start,
 373                                             (unsigned long)__hyp_text_end);
 374                 }
 375
 376                 swsusp_mte_restore_tags();
 377
 378                 /* make the crash dump kernel image protected again */
 379                 crash_post_resume();
 380
 381                 /*
 382                  * Tell the hibernation core that we've just restored
 383                  * the memory
 384                  */
 385                 in_suspend = 0;
 386
 387                 sleep_cpu = -EINVAL;
 388                 __cpu_suspend_exit();
 389
 390                 /*
 391                  * Just in case the boot kernel did turn the SSBD
 392                  * mitigation off behind our back, let's set the state
 393                  * to what we expect it to be.
 394                  */
 395                 spectre_v4_enable_mitigation(NULL);
 396         }
 397
 398         local_daif_restore(flags);
 399
 400         return ret;
 401 }
 402
 403 /*
 404  * Setup then Resume from the hibernate image using swsusp_arch_suspend_exit().
 405  *
 406  * Memory allocated by get_safe_page() will be dealt with by the hibernate code,
 407  * we don't need to free it here.
 408  */
 409 int swsusp_arch_resume(void)
 410 {
 411         int rc;
 412         void *zero_page;
 413         size_t exit_size;
 414         pgd_t *tmp_pg_dir;
 415         phys_addr_t el2_vectors;
 416         void __noreturn (*hibernate_exit)(phys_addr_t, phys_addr_t, void *,
 417                                           void *, phys_addr_t, phys_addr_t);
 418         struct trans_pgd_info trans_info = {
 419                 .trans_alloc_page       = hibernate_page_alloc,
 420                 .trans_alloc_arg        = (void *)GFP_ATOMIC,
 421         };
 422
 423         /*
 424          * Restoring the memory image will overwrite the ttbr1 page tables.
 425          * Create a second copy of just the linear map, and use this when
 426          * restoring.
 427          */
 428         rc = trans_pgd_create_copy(&trans_info, &tmp_pg_dir, PAGE_OFFSET,
 429                                    PAGE_END);
 430         if (rc)
 431                 return rc;
 432
 433         /*
 434          * We need a zero page that is zero before & after resume in order to
 435          * to break before make on the ttbr1 page tables.
 436          */
 437         zero_page = (void *)get_safe_page(GFP_ATOMIC);
 438         if (!zero_page) {
 439                 pr_err("Failed to allocate zero page.\n");
 440                 return -ENOMEM;
 441         }
 442
 443         if (el2_reset_needed()) {
 444                 rc = trans_pgd_copy_el2_vectors(&trans_info, &el2_vectors);
 445                 if (rc) {
 446                         pr_err("Failed to setup el2 vectors\n");
 447                         return rc;
 448                 }
 449         }
 450
 451         exit_size = __hibernate_exit_text_end - __hibernate_exit_text_start;
 452         /*
 453          * Copy swsusp_arch_suspend_exit() to a safe page. This will generate
 454          * a new set of ttbr0 page tables and load them.
 455          */
 456         rc = create_safe_exec_page(__hibernate_exit_text_start, exit_size,
 457                                    (phys_addr_t *)&hibernate_exit);
 458         if (rc) {
 459                 pr_err("Failed to create safe executable page for hibernate_exit code.\n");
 460                 return rc;
 461         }
 462
 463         /*
 464          * KASLR will cause the el2 vectors to be in a different location in
 465          * the resumed kernel. Load hibernate's temporary copy into el2.
 466          *
 467          * We can skip this step if we booted at EL1, or are running with VHE.
 468          */
 469         if (el2_reset_needed())
 470                 __hyp_set_vectors(el2_vectors);
 471
 472         hibernate_exit(virt_to_phys(tmp_pg_dir), resume_hdr.ttbr1_el1,
 473                        resume_hdr.reenter_kernel, restore_pblist,
 474                        resume_hdr.__hyp_stub_vectors, virt_to_phys(zero_page));
 475
 476         return 0;
 477 }
 478
 479 int hibernate_resume_nonboot_cpu_disable(void)
 480 {
 481         if (sleep_cpu < 0) {
 482                 pr_err("Failing to resume from hibernate on an unknown CPU.\n");
 483                 return -ENODEV;
 484         }
 485
 486         return freeze_secondary_cpus(sleep_cpu);
 487 }