Merge tag 'locking-urgent-2021-05-09' of git://git.kernel.org/pub/scm/linux/kernel...
[linux-2.6-microblaze.git] / 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_mode_available() && !is_kernel_in_hyp_mode())
53
54 /* temporary el2 vectors in the __hibernate_exit_text section. */
55 extern char hibernate_el2_vectors[];
56
57 /* hyp-stub vectors, used to restore el2 during resume from hibernate. */
58 extern char __hyp_stub_vectors[];
59
60 /*
61  * The logical cpu number we should resume on, initialised to a non-cpu
62  * number.
63  */
64 static int sleep_cpu = -EINVAL;
65
66 /*
67  * Values that may not change over hibernate/resume. We put the build number
68  * and date in here so that we guarantee not to resume with a different
69  * kernel.
70  */
71 struct arch_hibernate_hdr_invariants {
72         char            uts_version[__NEW_UTS_LEN + 1];
73 };
74
75 /* These values need to be know across a hibernate/restore. */
76 static struct arch_hibernate_hdr {
77         struct arch_hibernate_hdr_invariants invariants;
78
79         /* These are needed to find the relocated kernel if built with kaslr */
80         phys_addr_t     ttbr1_el1;
81         void            (*reenter_kernel)(void);
82
83         /*
84          * We need to know where the __hyp_stub_vectors are after restore to
85          * re-configure el2.
86          */
87         phys_addr_t     __hyp_stub_vectors;
88
89         u64             sleep_cpu_mpidr;
90 } resume_hdr;
91
92 static inline void arch_hdr_invariants(struct arch_hibernate_hdr_invariants *i)
93 {
94         memset(i, 0, sizeof(*i));
95         memcpy(i->uts_version, init_utsname()->version, sizeof(i->uts_version));
96 }
97
98 int pfn_is_nosave(unsigned long pfn)
99 {
100         unsigned long nosave_begin_pfn = sym_to_pfn(&__nosave_begin);
101         unsigned long nosave_end_pfn = sym_to_pfn(&__nosave_end - 1);
102
103         return ((pfn >= nosave_begin_pfn) && (pfn <= nosave_end_pfn)) ||
104                 crash_is_nosave(pfn);
105 }
106
107 void notrace save_processor_state(void)
108 {
109         WARN_ON(num_online_cpus() != 1);
110 }
111
112 void notrace restore_processor_state(void)
113 {
114 }
115
116 int arch_hibernation_header_save(void *addr, unsigned int max_size)
117 {
118         struct arch_hibernate_hdr *hdr = addr;
119
120         if (max_size < sizeof(*hdr))
121                 return -EOVERFLOW;
122
123         arch_hdr_invariants(&hdr->invariants);
124         hdr->ttbr1_el1          = __pa_symbol(swapper_pg_dir);
125         hdr->reenter_kernel     = _cpu_resume;
126
127         /* We can't use __hyp_get_vectors() because kvm may still be loaded */
128         if (el2_reset_needed())
129                 hdr->__hyp_stub_vectors = __pa_symbol(__hyp_stub_vectors);
130         else
131                 hdr->__hyp_stub_vectors = 0;
132
133         /* Save the mpidr of the cpu we called cpu_suspend() on... */
134         if (sleep_cpu < 0) {
135                 pr_err("Failing to hibernate on an unknown CPU.\n");
136                 return -ENODEV;
137         }
138         hdr->sleep_cpu_mpidr = cpu_logical_map(sleep_cpu);
139         pr_info("Hibernating on CPU %d [mpidr:0x%llx]\n", sleep_cpu,
140                 hdr->sleep_cpu_mpidr);
141
142         return 0;
143 }
144 EXPORT_SYMBOL(arch_hibernation_header_save);
145
146 int arch_hibernation_header_restore(void *addr)
147 {
148         int ret;
149         struct arch_hibernate_hdr_invariants invariants;
150         struct arch_hibernate_hdr *hdr = addr;
151
152         arch_hdr_invariants(&invariants);
153         if (memcmp(&hdr->invariants, &invariants, sizeof(invariants))) {
154                 pr_crit("Hibernate image not generated by this kernel!\n");
155                 return -EINVAL;
156         }
157
158         sleep_cpu = get_logical_index(hdr->sleep_cpu_mpidr);
159         pr_info("Hibernated on CPU %d [mpidr:0x%llx]\n", sleep_cpu,
160                 hdr->sleep_cpu_mpidr);
161         if (sleep_cpu < 0) {
162                 pr_crit("Hibernated on a CPU not known to this kernel!\n");
163                 sleep_cpu = -EINVAL;
164                 return -EINVAL;
165         }
166
167         ret = bringup_hibernate_cpu(sleep_cpu);
168         if (ret) {
169                 sleep_cpu = -EINVAL;
170                 return ret;
171         }
172
173         resume_hdr = *hdr;
174
175         return 0;
176 }
177 EXPORT_SYMBOL(arch_hibernation_header_restore);
178
179 static void *hibernate_page_alloc(void *arg)
180 {
181         return (void *)get_safe_page((__force gfp_t)(unsigned long)arg);
182 }
183
184 /*
185  * Copies length bytes, starting at src_start into an new page,
186  * perform cache maintenance, then maps it at the specified address low
187  * address as executable.
188  *
189  * This is used by hibernate to copy the code it needs to execute when
190  * overwriting the kernel text. This function generates a new set of page
191  * tables, which it loads into ttbr0.
192  *
193  * Length is provided as we probably only want 4K of data, even on a 64K
194  * page system.
195  */
196 static int create_safe_exec_page(void *src_start, size_t length,
197                                  phys_addr_t *phys_dst_addr)
198 {
199         struct trans_pgd_info trans_info = {
200                 .trans_alloc_page       = hibernate_page_alloc,
201                 .trans_alloc_arg        = (__force void *)GFP_ATOMIC,
202         };
203
204         void *page = (void *)get_safe_page(GFP_ATOMIC);
205         phys_addr_t trans_ttbr0;
206         unsigned long t0sz;
207         int rc;
208
209         if (!page)
210                 return -ENOMEM;
211
212         memcpy(page, src_start, length);
213         __flush_icache_range((unsigned long)page, (unsigned long)page + length);
214         rc = trans_pgd_idmap_page(&trans_info, &trans_ttbr0, &t0sz, page);
215         if (rc)
216                 return rc;
217
218         /*
219          * Load our new page tables. A strict BBM approach requires that we
220          * ensure that TLBs are free of any entries that may overlap with the
221          * global mappings we are about to install.
222          *
223          * For a real hibernate/resume cycle TTBR0 currently points to a zero
224          * page, but TLBs may contain stale ASID-tagged entries (e.g. for EFI
225          * runtime services), while for a userspace-driven test_resume cycle it
226          * points to userspace page tables (and we must point it at a zero page
227          * ourselves).
228          *
229          * We change T0SZ as part of installing the idmap. This is undone by
230          * cpu_uninstall_idmap() in __cpu_suspend_exit().
231          */
232         cpu_set_reserved_ttbr0();
233         local_flush_tlb_all();
234         __cpu_set_tcr_t0sz(t0sz);
235         write_sysreg(trans_ttbr0, ttbr0_el1);
236         isb();
237
238         *phys_dst_addr = virt_to_phys(page);
239
240         return 0;
241 }
242
243 #define dcache_clean_range(start, end)  __flush_dcache_area(start, (end - start))
244
245 #ifdef CONFIG_ARM64_MTE
246
247 static DEFINE_XARRAY(mte_pages);
248
249 static int save_tags(struct page *page, unsigned long pfn)
250 {
251         void *tag_storage, *ret;
252
253         tag_storage = mte_allocate_tag_storage();
254         if (!tag_storage)
255                 return -ENOMEM;
256
257         mte_save_page_tags(page_address(page), tag_storage);
258
259         ret = xa_store(&mte_pages, pfn, tag_storage, GFP_KERNEL);
260         if (WARN(xa_is_err(ret), "Failed to store MTE tags")) {
261                 mte_free_tag_storage(tag_storage);
262                 return xa_err(ret);
263         } else if (WARN(ret, "swsusp: %s: Duplicate entry", __func__)) {
264                 mte_free_tag_storage(ret);
265         }
266
267         return 0;
268 }
269
270 static void swsusp_mte_free_storage(void)
271 {
272         XA_STATE(xa_state, &mte_pages, 0);
273         void *tags;
274
275         xa_lock(&mte_pages);
276         xas_for_each(&xa_state, tags, ULONG_MAX) {
277                 mte_free_tag_storage(tags);
278         }
279         xa_unlock(&mte_pages);
280
281         xa_destroy(&mte_pages);
282 }
283
284 static int swsusp_mte_save_tags(void)
285 {
286         struct zone *zone;
287         unsigned long pfn, max_zone_pfn;
288         int ret = 0;
289         int n = 0;
290
291         if (!system_supports_mte())
292                 return 0;
293
294         for_each_populated_zone(zone) {
295                 max_zone_pfn = zone_end_pfn(zone);
296                 for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++) {
297                         struct page *page = pfn_to_online_page(pfn);
298
299                         if (!page)
300                                 continue;
301
302                         if (!test_bit(PG_mte_tagged, &page->flags))
303                                 continue;
304
305                         ret = save_tags(page, pfn);
306                         if (ret) {
307                                 swsusp_mte_free_storage();
308                                 goto out;
309                         }
310
311                         n++;
312                 }
313         }
314         pr_info("Saved %d MTE pages\n", n);
315
316 out:
317         return ret;
318 }
319
320 static void swsusp_mte_restore_tags(void)
321 {
322         XA_STATE(xa_state, &mte_pages, 0);
323         int n = 0;
324         void *tags;
325
326         xa_lock(&mte_pages);
327         xas_for_each(&xa_state, tags, ULONG_MAX) {
328                 unsigned long pfn = xa_state.xa_index;
329                 struct page *page = pfn_to_online_page(pfn);
330
331                 /*
332                  * It is not required to invoke page_kasan_tag_reset(page)
333                  * at this point since the tags stored in page->flags are
334                  * already restored.
335                  */
336                 mte_restore_page_tags(page_address(page), tags);
337
338                 mte_free_tag_storage(tags);
339                 n++;
340         }
341         xa_unlock(&mte_pages);
342
343         pr_info("Restored %d MTE pages\n", n);
344
345         xa_destroy(&mte_pages);
346 }
347
348 #else   /* CONFIG_ARM64_MTE */
349
350 static int swsusp_mte_save_tags(void)
351 {
352         return 0;
353 }
354
355 static void swsusp_mte_restore_tags(void)
356 {
357 }
358
359 #endif  /* CONFIG_ARM64_MTE */
360
361 int swsusp_arch_suspend(void)
362 {
363         int ret = 0;
364         unsigned long flags;
365         struct sleep_stack_data state;
366
367         if (cpus_are_stuck_in_kernel()) {
368                 pr_err("Can't hibernate: no mechanism to offline secondary CPUs.\n");
369                 return -EBUSY;
370         }
371
372         flags = local_daif_save();
373
374         if (__cpu_suspend_enter(&state)) {
375                 /* make the crash dump kernel image visible/saveable */
376                 crash_prepare_suspend();
377
378                 ret = swsusp_mte_save_tags();
379                 if (ret)
380                         return ret;
381
382                 sleep_cpu = smp_processor_id();
383                 ret = swsusp_save();
384         } else {
385                 /* Clean kernel core startup/idle code to PoC*/
386                 dcache_clean_range(__mmuoff_data_start, __mmuoff_data_end);
387                 dcache_clean_range(__idmap_text_start, __idmap_text_end);
388
389                 /* Clean kvm setup code to PoC? */
390                 if (el2_reset_needed()) {
391                         dcache_clean_range(__hyp_idmap_text_start, __hyp_idmap_text_end);
392                         dcache_clean_range(__hyp_text_start, __hyp_text_end);
393                 }
394
395                 swsusp_mte_restore_tags();
396
397                 /* make the crash dump kernel image protected again */
398                 crash_post_resume();
399
400                 /*
401                  * Tell the hibernation core that we've just restored
402                  * the memory
403                  */
404                 in_suspend = 0;
405
406                 sleep_cpu = -EINVAL;
407                 __cpu_suspend_exit();
408
409                 /*
410                  * Just in case the boot kernel did turn the SSBD
411                  * mitigation off behind our back, let's set the state
412                  * to what we expect it to be.
413                  */
414                 spectre_v4_enable_mitigation(NULL);
415         }
416
417         local_daif_restore(flags);
418
419         return ret;
420 }
421
422 /*
423  * Setup then Resume from the hibernate image using swsusp_arch_suspend_exit().
424  *
425  * Memory allocated by get_safe_page() will be dealt with by the hibernate code,
426  * we don't need to free it here.
427  */
428 int swsusp_arch_resume(void)
429 {
430         int rc;
431         void *zero_page;
432         size_t exit_size;
433         pgd_t *tmp_pg_dir;
434         void __noreturn (*hibernate_exit)(phys_addr_t, phys_addr_t, void *,
435                                           void *, phys_addr_t, phys_addr_t);
436         struct trans_pgd_info trans_info = {
437                 .trans_alloc_page       = hibernate_page_alloc,
438                 .trans_alloc_arg        = (void *)GFP_ATOMIC,
439         };
440
441         /*
442          * Restoring the memory image will overwrite the ttbr1 page tables.
443          * Create a second copy of just the linear map, and use this when
444          * restoring.
445          */
446         rc = trans_pgd_create_copy(&trans_info, &tmp_pg_dir, PAGE_OFFSET,
447                                    PAGE_END);
448         if (rc)
449                 return rc;
450
451         /*
452          * We need a zero page that is zero before & after resume in order to
453          * to break before make on the ttbr1 page tables.
454          */
455         zero_page = (void *)get_safe_page(GFP_ATOMIC);
456         if (!zero_page) {
457                 pr_err("Failed to allocate zero page.\n");
458                 return -ENOMEM;
459         }
460
461         exit_size = __hibernate_exit_text_end - __hibernate_exit_text_start;
462         /*
463          * Copy swsusp_arch_suspend_exit() to a safe page. This will generate
464          * a new set of ttbr0 page tables and load them.
465          */
466         rc = create_safe_exec_page(__hibernate_exit_text_start, exit_size,
467                                    (phys_addr_t *)&hibernate_exit);
468         if (rc) {
469                 pr_err("Failed to create safe executable page for hibernate_exit code.\n");
470                 return rc;
471         }
472
473         /*
474          * The hibernate exit text contains a set of el2 vectors, that will
475          * be executed at el2 with the mmu off in order to reload hyp-stub.
476          */
477         __flush_dcache_area(hibernate_exit, exit_size);
478
479         /*
480          * KASLR will cause the el2 vectors to be in a different location in
481          * the resumed kernel. Load hibernate's temporary copy into el2.
482          *
483          * We can skip this step if we booted at EL1, or are running with VHE.
484          */
485         if (el2_reset_needed()) {
486                 phys_addr_t el2_vectors = (phys_addr_t)hibernate_exit;
487                 el2_vectors += hibernate_el2_vectors -
488                                __hibernate_exit_text_start;     /* offset */
489
490                 __hyp_set_vectors(el2_vectors);
491         }
492
493         hibernate_exit(virt_to_phys(tmp_pg_dir), resume_hdr.ttbr1_el1,
494                        resume_hdr.reenter_kernel, restore_pblist,
495                        resume_hdr.__hyp_stub_vectors, virt_to_phys(zero_page));
496
497         return 0;
498 }
499
500 int hibernate_resume_nonboot_cpu_disable(void)
501 {
502         if (sleep_cpu < 0) {
503                 pr_err("Failing to resume from hibernate on an unknown CPU.\n");
504                 return -ENODEV;
505         }
506
507         return freeze_secondary_cpus(sleep_cpu);
508 }