Loading...
1/*
2 * kexec.c - kexec system call core code.
3 * Copyright (C) 2002-2004 Eric Biederman <ebiederm@xmission.com>
4 *
5 * This source code is licensed under the GNU General Public License,
6 * Version 2. See the file COPYING for more details.
7 */
8
9#define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
10
11#include <linux/capability.h>
12#include <linux/mm.h>
13#include <linux/file.h>
14#include <linux/slab.h>
15#include <linux/fs.h>
16#include <linux/kexec.h>
17#include <linux/mutex.h>
18#include <linux/list.h>
19#include <linux/highmem.h>
20#include <linux/syscalls.h>
21#include <linux/reboot.h>
22#include <linux/ioport.h>
23#include <linux/hardirq.h>
24#include <linux/elf.h>
25#include <linux/elfcore.h>
26#include <linux/utsname.h>
27#include <linux/numa.h>
28#include <linux/suspend.h>
29#include <linux/device.h>
30#include <linux/freezer.h>
31#include <linux/pm.h>
32#include <linux/cpu.h>
33#include <linux/uaccess.h>
34#include <linux/io.h>
35#include <linux/console.h>
36#include <linux/vmalloc.h>
37#include <linux/swap.h>
38#include <linux/syscore_ops.h>
39#include <linux/compiler.h>
40#include <linux/hugetlb.h>
41#include <linux/frame.h>
42
43#include <asm/page.h>
44#include <asm/sections.h>
45
46#include <crypto/hash.h>
47#include <crypto/sha.h>
48#include "kexec_internal.h"
49
50DEFINE_MUTEX(kexec_mutex);
51
52/* Per cpu memory for storing cpu states in case of system crash. */
53note_buf_t __percpu *crash_notes;
54
55/* Flag to indicate we are going to kexec a new kernel */
56bool kexec_in_progress = false;
57
58
59/* Location of the reserved area for the crash kernel */
60struct resource crashk_res = {
61 .name = "Crash kernel",
62 .start = 0,
63 .end = 0,
64 .flags = IORESOURCE_BUSY | IORESOURCE_SYSTEM_RAM,
65 .desc = IORES_DESC_CRASH_KERNEL
66};
67struct resource crashk_low_res = {
68 .name = "Crash kernel",
69 .start = 0,
70 .end = 0,
71 .flags = IORESOURCE_BUSY | IORESOURCE_SYSTEM_RAM,
72 .desc = IORES_DESC_CRASH_KERNEL
73};
74
75int kexec_should_crash(struct task_struct *p)
76{
77 /*
78 * If crash_kexec_post_notifiers is enabled, don't run
79 * crash_kexec() here yet, which must be run after panic
80 * notifiers in panic().
81 */
82 if (crash_kexec_post_notifiers)
83 return 0;
84 /*
85 * There are 4 panic() calls in do_exit() path, each of which
86 * corresponds to each of these 4 conditions.
87 */
88 if (in_interrupt() || !p->pid || is_global_init(p) || panic_on_oops)
89 return 1;
90 return 0;
91}
92
93int kexec_crash_loaded(void)
94{
95 return !!kexec_crash_image;
96}
97EXPORT_SYMBOL_GPL(kexec_crash_loaded);
98
99/*
100 * When kexec transitions to the new kernel there is a one-to-one
101 * mapping between physical and virtual addresses. On processors
102 * where you can disable the MMU this is trivial, and easy. For
103 * others it is still a simple predictable page table to setup.
104 *
105 * In that environment kexec copies the new kernel to its final
106 * resting place. This means I can only support memory whose
107 * physical address can fit in an unsigned long. In particular
108 * addresses where (pfn << PAGE_SHIFT) > ULONG_MAX cannot be handled.
109 * If the assembly stub has more restrictive requirements
110 * KEXEC_SOURCE_MEMORY_LIMIT and KEXEC_DEST_MEMORY_LIMIT can be
111 * defined more restrictively in <asm/kexec.h>.
112 *
113 * The code for the transition from the current kernel to the
114 * the new kernel is placed in the control_code_buffer, whose size
115 * is given by KEXEC_CONTROL_PAGE_SIZE. In the best case only a single
116 * page of memory is necessary, but some architectures require more.
117 * Because this memory must be identity mapped in the transition from
118 * virtual to physical addresses it must live in the range
119 * 0 - TASK_SIZE, as only the user space mappings are arbitrarily
120 * modifiable.
121 *
122 * The assembly stub in the control code buffer is passed a linked list
123 * of descriptor pages detailing the source pages of the new kernel,
124 * and the destination addresses of those source pages. As this data
125 * structure is not used in the context of the current OS, it must
126 * be self-contained.
127 *
128 * The code has been made to work with highmem pages and will use a
129 * destination page in its final resting place (if it happens
130 * to allocate it). The end product of this is that most of the
131 * physical address space, and most of RAM can be used.
132 *
133 * Future directions include:
134 * - allocating a page table with the control code buffer identity
135 * mapped, to simplify machine_kexec and make kexec_on_panic more
136 * reliable.
137 */
138
139/*
140 * KIMAGE_NO_DEST is an impossible destination address..., for
141 * allocating pages whose destination address we do not care about.
142 */
143#define KIMAGE_NO_DEST (-1UL)
144#define PAGE_COUNT(x) (((x) + PAGE_SIZE - 1) >> PAGE_SHIFT)
145
146static struct page *kimage_alloc_page(struct kimage *image,
147 gfp_t gfp_mask,
148 unsigned long dest);
149
150int sanity_check_segment_list(struct kimage *image)
151{
152 int i;
153 unsigned long nr_segments = image->nr_segments;
154 unsigned long total_pages = 0;
155
156 /*
157 * Verify we have good destination addresses. The caller is
158 * responsible for making certain we don't attempt to load
159 * the new image into invalid or reserved areas of RAM. This
160 * just verifies it is an address we can use.
161 *
162 * Since the kernel does everything in page size chunks ensure
163 * the destination addresses are page aligned. Too many
164 * special cases crop of when we don't do this. The most
165 * insidious is getting overlapping destination addresses
166 * simply because addresses are changed to page size
167 * granularity.
168 */
169 for (i = 0; i < nr_segments; i++) {
170 unsigned long mstart, mend;
171
172 mstart = image->segment[i].mem;
173 mend = mstart + image->segment[i].memsz;
174 if (mstart > mend)
175 return -EADDRNOTAVAIL;
176 if ((mstart & ~PAGE_MASK) || (mend & ~PAGE_MASK))
177 return -EADDRNOTAVAIL;
178 if (mend >= KEXEC_DESTINATION_MEMORY_LIMIT)
179 return -EADDRNOTAVAIL;
180 }
181
182 /* Verify our destination addresses do not overlap.
183 * If we alloed overlapping destination addresses
184 * through very weird things can happen with no
185 * easy explanation as one segment stops on another.
186 */
187 for (i = 0; i < nr_segments; i++) {
188 unsigned long mstart, mend;
189 unsigned long j;
190
191 mstart = image->segment[i].mem;
192 mend = mstart + image->segment[i].memsz;
193 for (j = 0; j < i; j++) {
194 unsigned long pstart, pend;
195
196 pstart = image->segment[j].mem;
197 pend = pstart + image->segment[j].memsz;
198 /* Do the segments overlap ? */
199 if ((mend > pstart) && (mstart < pend))
200 return -EINVAL;
201 }
202 }
203
204 /* Ensure our buffer sizes are strictly less than
205 * our memory sizes. This should always be the case,
206 * and it is easier to check up front than to be surprised
207 * later on.
208 */
209 for (i = 0; i < nr_segments; i++) {
210 if (image->segment[i].bufsz > image->segment[i].memsz)
211 return -EINVAL;
212 }
213
214 /*
215 * Verify that no more than half of memory will be consumed. If the
216 * request from userspace is too large, a large amount of time will be
217 * wasted allocating pages, which can cause a soft lockup.
218 */
219 for (i = 0; i < nr_segments; i++) {
220 if (PAGE_COUNT(image->segment[i].memsz) > totalram_pages / 2)
221 return -EINVAL;
222
223 total_pages += PAGE_COUNT(image->segment[i].memsz);
224 }
225
226 if (total_pages > totalram_pages / 2)
227 return -EINVAL;
228
229 /*
230 * Verify we have good destination addresses. Normally
231 * the caller is responsible for making certain we don't
232 * attempt to load the new image into invalid or reserved
233 * areas of RAM. But crash kernels are preloaded into a
234 * reserved area of ram. We must ensure the addresses
235 * are in the reserved area otherwise preloading the
236 * kernel could corrupt things.
237 */
238
239 if (image->type == KEXEC_TYPE_CRASH) {
240 for (i = 0; i < nr_segments; i++) {
241 unsigned long mstart, mend;
242
243 mstart = image->segment[i].mem;
244 mend = mstart + image->segment[i].memsz - 1;
245 /* Ensure we are within the crash kernel limits */
246 if ((mstart < phys_to_boot_phys(crashk_res.start)) ||
247 (mend > phys_to_boot_phys(crashk_res.end)))
248 return -EADDRNOTAVAIL;
249 }
250 }
251
252 return 0;
253}
254
255struct kimage *do_kimage_alloc_init(void)
256{
257 struct kimage *image;
258
259 /* Allocate a controlling structure */
260 image = kzalloc(sizeof(*image), GFP_KERNEL);
261 if (!image)
262 return NULL;
263
264 image->head = 0;
265 image->entry = &image->head;
266 image->last_entry = &image->head;
267 image->control_page = ~0; /* By default this does not apply */
268 image->type = KEXEC_TYPE_DEFAULT;
269
270 /* Initialize the list of control pages */
271 INIT_LIST_HEAD(&image->control_pages);
272
273 /* Initialize the list of destination pages */
274 INIT_LIST_HEAD(&image->dest_pages);
275
276 /* Initialize the list of unusable pages */
277 INIT_LIST_HEAD(&image->unusable_pages);
278
279 return image;
280}
281
282int kimage_is_destination_range(struct kimage *image,
283 unsigned long start,
284 unsigned long end)
285{
286 unsigned long i;
287
288 for (i = 0; i < image->nr_segments; i++) {
289 unsigned long mstart, mend;
290
291 mstart = image->segment[i].mem;
292 mend = mstart + image->segment[i].memsz;
293 if ((end > mstart) && (start < mend))
294 return 1;
295 }
296
297 return 0;
298}
299
300static struct page *kimage_alloc_pages(gfp_t gfp_mask, unsigned int order)
301{
302 struct page *pages;
303
304 pages = alloc_pages(gfp_mask & ~__GFP_ZERO, order);
305 if (pages) {
306 unsigned int count, i;
307
308 pages->mapping = NULL;
309 set_page_private(pages, order);
310 count = 1 << order;
311 for (i = 0; i < count; i++)
312 SetPageReserved(pages + i);
313
314 arch_kexec_post_alloc_pages(page_address(pages), count,
315 gfp_mask);
316
317 if (gfp_mask & __GFP_ZERO)
318 for (i = 0; i < count; i++)
319 clear_highpage(pages + i);
320 }
321
322 return pages;
323}
324
325static void kimage_free_pages(struct page *page)
326{
327 unsigned int order, count, i;
328
329 order = page_private(page);
330 count = 1 << order;
331
332 arch_kexec_pre_free_pages(page_address(page), count);
333
334 for (i = 0; i < count; i++)
335 ClearPageReserved(page + i);
336 __free_pages(page, order);
337}
338
339void kimage_free_page_list(struct list_head *list)
340{
341 struct page *page, *next;
342
343 list_for_each_entry_safe(page, next, list, lru) {
344 list_del(&page->lru);
345 kimage_free_pages(page);
346 }
347}
348
349static struct page *kimage_alloc_normal_control_pages(struct kimage *image,
350 unsigned int order)
351{
352 /* Control pages are special, they are the intermediaries
353 * that are needed while we copy the rest of the pages
354 * to their final resting place. As such they must
355 * not conflict with either the destination addresses
356 * or memory the kernel is already using.
357 *
358 * The only case where we really need more than one of
359 * these are for architectures where we cannot disable
360 * the MMU and must instead generate an identity mapped
361 * page table for all of the memory.
362 *
363 * At worst this runs in O(N) of the image size.
364 */
365 struct list_head extra_pages;
366 struct page *pages;
367 unsigned int count;
368
369 count = 1 << order;
370 INIT_LIST_HEAD(&extra_pages);
371
372 /* Loop while I can allocate a page and the page allocated
373 * is a destination page.
374 */
375 do {
376 unsigned long pfn, epfn, addr, eaddr;
377
378 pages = kimage_alloc_pages(KEXEC_CONTROL_MEMORY_GFP, order);
379 if (!pages)
380 break;
381 pfn = page_to_boot_pfn(pages);
382 epfn = pfn + count;
383 addr = pfn << PAGE_SHIFT;
384 eaddr = epfn << PAGE_SHIFT;
385 if ((epfn >= (KEXEC_CONTROL_MEMORY_LIMIT >> PAGE_SHIFT)) ||
386 kimage_is_destination_range(image, addr, eaddr)) {
387 list_add(&pages->lru, &extra_pages);
388 pages = NULL;
389 }
390 } while (!pages);
391
392 if (pages) {
393 /* Remember the allocated page... */
394 list_add(&pages->lru, &image->control_pages);
395
396 /* Because the page is already in it's destination
397 * location we will never allocate another page at
398 * that address. Therefore kimage_alloc_pages
399 * will not return it (again) and we don't need
400 * to give it an entry in image->segment[].
401 */
402 }
403 /* Deal with the destination pages I have inadvertently allocated.
404 *
405 * Ideally I would convert multi-page allocations into single
406 * page allocations, and add everything to image->dest_pages.
407 *
408 * For now it is simpler to just free the pages.
409 */
410 kimage_free_page_list(&extra_pages);
411
412 return pages;
413}
414
415static struct page *kimage_alloc_crash_control_pages(struct kimage *image,
416 unsigned int order)
417{
418 /* Control pages are special, they are the intermediaries
419 * that are needed while we copy the rest of the pages
420 * to their final resting place. As such they must
421 * not conflict with either the destination addresses
422 * or memory the kernel is already using.
423 *
424 * Control pages are also the only pags we must allocate
425 * when loading a crash kernel. All of the other pages
426 * are specified by the segments and we just memcpy
427 * into them directly.
428 *
429 * The only case where we really need more than one of
430 * these are for architectures where we cannot disable
431 * the MMU and must instead generate an identity mapped
432 * page table for all of the memory.
433 *
434 * Given the low demand this implements a very simple
435 * allocator that finds the first hole of the appropriate
436 * size in the reserved memory region, and allocates all
437 * of the memory up to and including the hole.
438 */
439 unsigned long hole_start, hole_end, size;
440 struct page *pages;
441
442 pages = NULL;
443 size = (1 << order) << PAGE_SHIFT;
444 hole_start = (image->control_page + (size - 1)) & ~(size - 1);
445 hole_end = hole_start + size - 1;
446 while (hole_end <= crashk_res.end) {
447 unsigned long i;
448
449 cond_resched();
450
451 if (hole_end > KEXEC_CRASH_CONTROL_MEMORY_LIMIT)
452 break;
453 /* See if I overlap any of the segments */
454 for (i = 0; i < image->nr_segments; i++) {
455 unsigned long mstart, mend;
456
457 mstart = image->segment[i].mem;
458 mend = mstart + image->segment[i].memsz - 1;
459 if ((hole_end >= mstart) && (hole_start <= mend)) {
460 /* Advance the hole to the end of the segment */
461 hole_start = (mend + (size - 1)) & ~(size - 1);
462 hole_end = hole_start + size - 1;
463 break;
464 }
465 }
466 /* If I don't overlap any segments I have found my hole! */
467 if (i == image->nr_segments) {
468 pages = pfn_to_page(hole_start >> PAGE_SHIFT);
469 image->control_page = hole_end;
470 break;
471 }
472 }
473
474 return pages;
475}
476
477
478struct page *kimage_alloc_control_pages(struct kimage *image,
479 unsigned int order)
480{
481 struct page *pages = NULL;
482
483 switch (image->type) {
484 case KEXEC_TYPE_DEFAULT:
485 pages = kimage_alloc_normal_control_pages(image, order);
486 break;
487 case KEXEC_TYPE_CRASH:
488 pages = kimage_alloc_crash_control_pages(image, order);
489 break;
490 }
491
492 return pages;
493}
494
495int kimage_crash_copy_vmcoreinfo(struct kimage *image)
496{
497 struct page *vmcoreinfo_page;
498 void *safecopy;
499
500 if (image->type != KEXEC_TYPE_CRASH)
501 return 0;
502
503 /*
504 * For kdump, allocate one vmcoreinfo safe copy from the
505 * crash memory. as we have arch_kexec_protect_crashkres()
506 * after kexec syscall, we naturally protect it from write
507 * (even read) access under kernel direct mapping. But on
508 * the other hand, we still need to operate it when crash
509 * happens to generate vmcoreinfo note, hereby we rely on
510 * vmap for this purpose.
511 */
512 vmcoreinfo_page = kimage_alloc_control_pages(image, 0);
513 if (!vmcoreinfo_page) {
514 pr_warn("Could not allocate vmcoreinfo buffer\n");
515 return -ENOMEM;
516 }
517 safecopy = vmap(&vmcoreinfo_page, 1, VM_MAP, PAGE_KERNEL);
518 if (!safecopy) {
519 pr_warn("Could not vmap vmcoreinfo buffer\n");
520 return -ENOMEM;
521 }
522
523 image->vmcoreinfo_data_copy = safecopy;
524 crash_update_vmcoreinfo_safecopy(safecopy);
525
526 return 0;
527}
528
529static int kimage_add_entry(struct kimage *image, kimage_entry_t entry)
530{
531 if (*image->entry != 0)
532 image->entry++;
533
534 if (image->entry == image->last_entry) {
535 kimage_entry_t *ind_page;
536 struct page *page;
537
538 page = kimage_alloc_page(image, GFP_KERNEL, KIMAGE_NO_DEST);
539 if (!page)
540 return -ENOMEM;
541
542 ind_page = page_address(page);
543 *image->entry = virt_to_boot_phys(ind_page) | IND_INDIRECTION;
544 image->entry = ind_page;
545 image->last_entry = ind_page +
546 ((PAGE_SIZE/sizeof(kimage_entry_t)) - 1);
547 }
548 *image->entry = entry;
549 image->entry++;
550 *image->entry = 0;
551
552 return 0;
553}
554
555static int kimage_set_destination(struct kimage *image,
556 unsigned long destination)
557{
558 int result;
559
560 destination &= PAGE_MASK;
561 result = kimage_add_entry(image, destination | IND_DESTINATION);
562
563 return result;
564}
565
566
567static int kimage_add_page(struct kimage *image, unsigned long page)
568{
569 int result;
570
571 page &= PAGE_MASK;
572 result = kimage_add_entry(image, page | IND_SOURCE);
573
574 return result;
575}
576
577
578static void kimage_free_extra_pages(struct kimage *image)
579{
580 /* Walk through and free any extra destination pages I may have */
581 kimage_free_page_list(&image->dest_pages);
582
583 /* Walk through and free any unusable pages I have cached */
584 kimage_free_page_list(&image->unusable_pages);
585
586}
587void kimage_terminate(struct kimage *image)
588{
589 if (*image->entry != 0)
590 image->entry++;
591
592 *image->entry = IND_DONE;
593}
594
595#define for_each_kimage_entry(image, ptr, entry) \
596 for (ptr = &image->head; (entry = *ptr) && !(entry & IND_DONE); \
597 ptr = (entry & IND_INDIRECTION) ? \
598 boot_phys_to_virt((entry & PAGE_MASK)) : ptr + 1)
599
600static void kimage_free_entry(kimage_entry_t entry)
601{
602 struct page *page;
603
604 page = boot_pfn_to_page(entry >> PAGE_SHIFT);
605 kimage_free_pages(page);
606}
607
608void kimage_free(struct kimage *image)
609{
610 kimage_entry_t *ptr, entry;
611 kimage_entry_t ind = 0;
612
613 if (!image)
614 return;
615
616 if (image->vmcoreinfo_data_copy) {
617 crash_update_vmcoreinfo_safecopy(NULL);
618 vunmap(image->vmcoreinfo_data_copy);
619 }
620
621 kimage_free_extra_pages(image);
622 for_each_kimage_entry(image, ptr, entry) {
623 if (entry & IND_INDIRECTION) {
624 /* Free the previous indirection page */
625 if (ind & IND_INDIRECTION)
626 kimage_free_entry(ind);
627 /* Save this indirection page until we are
628 * done with it.
629 */
630 ind = entry;
631 } else if (entry & IND_SOURCE)
632 kimage_free_entry(entry);
633 }
634 /* Free the final indirection page */
635 if (ind & IND_INDIRECTION)
636 kimage_free_entry(ind);
637
638 /* Handle any machine specific cleanup */
639 machine_kexec_cleanup(image);
640
641 /* Free the kexec control pages... */
642 kimage_free_page_list(&image->control_pages);
643
644 /*
645 * Free up any temporary buffers allocated. This might hit if
646 * error occurred much later after buffer allocation.
647 */
648 if (image->file_mode)
649 kimage_file_post_load_cleanup(image);
650
651 kfree(image);
652}
653
654static kimage_entry_t *kimage_dst_used(struct kimage *image,
655 unsigned long page)
656{
657 kimage_entry_t *ptr, entry;
658 unsigned long destination = 0;
659
660 for_each_kimage_entry(image, ptr, entry) {
661 if (entry & IND_DESTINATION)
662 destination = entry & PAGE_MASK;
663 else if (entry & IND_SOURCE) {
664 if (page == destination)
665 return ptr;
666 destination += PAGE_SIZE;
667 }
668 }
669
670 return NULL;
671}
672
673static struct page *kimage_alloc_page(struct kimage *image,
674 gfp_t gfp_mask,
675 unsigned long destination)
676{
677 /*
678 * Here we implement safeguards to ensure that a source page
679 * is not copied to its destination page before the data on
680 * the destination page is no longer useful.
681 *
682 * To do this we maintain the invariant that a source page is
683 * either its own destination page, or it is not a
684 * destination page at all.
685 *
686 * That is slightly stronger than required, but the proof
687 * that no problems will not occur is trivial, and the
688 * implementation is simply to verify.
689 *
690 * When allocating all pages normally this algorithm will run
691 * in O(N) time, but in the worst case it will run in O(N^2)
692 * time. If the runtime is a problem the data structures can
693 * be fixed.
694 */
695 struct page *page;
696 unsigned long addr;
697
698 /*
699 * Walk through the list of destination pages, and see if I
700 * have a match.
701 */
702 list_for_each_entry(page, &image->dest_pages, lru) {
703 addr = page_to_boot_pfn(page) << PAGE_SHIFT;
704 if (addr == destination) {
705 list_del(&page->lru);
706 return page;
707 }
708 }
709 page = NULL;
710 while (1) {
711 kimage_entry_t *old;
712
713 /* Allocate a page, if we run out of memory give up */
714 page = kimage_alloc_pages(gfp_mask, 0);
715 if (!page)
716 return NULL;
717 /* If the page cannot be used file it away */
718 if (page_to_boot_pfn(page) >
719 (KEXEC_SOURCE_MEMORY_LIMIT >> PAGE_SHIFT)) {
720 list_add(&page->lru, &image->unusable_pages);
721 continue;
722 }
723 addr = page_to_boot_pfn(page) << PAGE_SHIFT;
724
725 /* If it is the destination page we want use it */
726 if (addr == destination)
727 break;
728
729 /* If the page is not a destination page use it */
730 if (!kimage_is_destination_range(image, addr,
731 addr + PAGE_SIZE))
732 break;
733
734 /*
735 * I know that the page is someones destination page.
736 * See if there is already a source page for this
737 * destination page. And if so swap the source pages.
738 */
739 old = kimage_dst_used(image, addr);
740 if (old) {
741 /* If so move it */
742 unsigned long old_addr;
743 struct page *old_page;
744
745 old_addr = *old & PAGE_MASK;
746 old_page = boot_pfn_to_page(old_addr >> PAGE_SHIFT);
747 copy_highpage(page, old_page);
748 *old = addr | (*old & ~PAGE_MASK);
749
750 /* The old page I have found cannot be a
751 * destination page, so return it if it's
752 * gfp_flags honor the ones passed in.
753 */
754 if (!(gfp_mask & __GFP_HIGHMEM) &&
755 PageHighMem(old_page)) {
756 kimage_free_pages(old_page);
757 continue;
758 }
759 addr = old_addr;
760 page = old_page;
761 break;
762 }
763 /* Place the page on the destination list, to be used later */
764 list_add(&page->lru, &image->dest_pages);
765 }
766
767 return page;
768}
769
770static int kimage_load_normal_segment(struct kimage *image,
771 struct kexec_segment *segment)
772{
773 unsigned long maddr;
774 size_t ubytes, mbytes;
775 int result;
776 unsigned char __user *buf = NULL;
777 unsigned char *kbuf = NULL;
778
779 result = 0;
780 if (image->file_mode)
781 kbuf = segment->kbuf;
782 else
783 buf = segment->buf;
784 ubytes = segment->bufsz;
785 mbytes = segment->memsz;
786 maddr = segment->mem;
787
788 result = kimage_set_destination(image, maddr);
789 if (result < 0)
790 goto out;
791
792 while (mbytes) {
793 struct page *page;
794 char *ptr;
795 size_t uchunk, mchunk;
796
797 page = kimage_alloc_page(image, GFP_HIGHUSER, maddr);
798 if (!page) {
799 result = -ENOMEM;
800 goto out;
801 }
802 result = kimage_add_page(image, page_to_boot_pfn(page)
803 << PAGE_SHIFT);
804 if (result < 0)
805 goto out;
806
807 ptr = kmap(page);
808 /* Start with a clear page */
809 clear_page(ptr);
810 ptr += maddr & ~PAGE_MASK;
811 mchunk = min_t(size_t, mbytes,
812 PAGE_SIZE - (maddr & ~PAGE_MASK));
813 uchunk = min(ubytes, mchunk);
814
815 /* For file based kexec, source pages are in kernel memory */
816 if (image->file_mode)
817 memcpy(ptr, kbuf, uchunk);
818 else
819 result = copy_from_user(ptr, buf, uchunk);
820 kunmap(page);
821 if (result) {
822 result = -EFAULT;
823 goto out;
824 }
825 ubytes -= uchunk;
826 maddr += mchunk;
827 if (image->file_mode)
828 kbuf += mchunk;
829 else
830 buf += mchunk;
831 mbytes -= mchunk;
832 }
833out:
834 return result;
835}
836
837static int kimage_load_crash_segment(struct kimage *image,
838 struct kexec_segment *segment)
839{
840 /* For crash dumps kernels we simply copy the data from
841 * user space to it's destination.
842 * We do things a page at a time for the sake of kmap.
843 */
844 unsigned long maddr;
845 size_t ubytes, mbytes;
846 int result;
847 unsigned char __user *buf = NULL;
848 unsigned char *kbuf = NULL;
849
850 result = 0;
851 if (image->file_mode)
852 kbuf = segment->kbuf;
853 else
854 buf = segment->buf;
855 ubytes = segment->bufsz;
856 mbytes = segment->memsz;
857 maddr = segment->mem;
858 while (mbytes) {
859 struct page *page;
860 char *ptr;
861 size_t uchunk, mchunk;
862
863 page = boot_pfn_to_page(maddr >> PAGE_SHIFT);
864 if (!page) {
865 result = -ENOMEM;
866 goto out;
867 }
868 ptr = kmap(page);
869 ptr += maddr & ~PAGE_MASK;
870 mchunk = min_t(size_t, mbytes,
871 PAGE_SIZE - (maddr & ~PAGE_MASK));
872 uchunk = min(ubytes, mchunk);
873 if (mchunk > uchunk) {
874 /* Zero the trailing part of the page */
875 memset(ptr + uchunk, 0, mchunk - uchunk);
876 }
877
878 /* For file based kexec, source pages are in kernel memory */
879 if (image->file_mode)
880 memcpy(ptr, kbuf, uchunk);
881 else
882 result = copy_from_user(ptr, buf, uchunk);
883 kexec_flush_icache_page(page);
884 kunmap(page);
885 if (result) {
886 result = -EFAULT;
887 goto out;
888 }
889 ubytes -= uchunk;
890 maddr += mchunk;
891 if (image->file_mode)
892 kbuf += mchunk;
893 else
894 buf += mchunk;
895 mbytes -= mchunk;
896 }
897out:
898 return result;
899}
900
901int kimage_load_segment(struct kimage *image,
902 struct kexec_segment *segment)
903{
904 int result = -ENOMEM;
905
906 switch (image->type) {
907 case KEXEC_TYPE_DEFAULT:
908 result = kimage_load_normal_segment(image, segment);
909 break;
910 case KEXEC_TYPE_CRASH:
911 result = kimage_load_crash_segment(image, segment);
912 break;
913 }
914
915 return result;
916}
917
918struct kimage *kexec_image;
919struct kimage *kexec_crash_image;
920int kexec_load_disabled;
921
922/*
923 * No panic_cpu check version of crash_kexec(). This function is called
924 * only when panic_cpu holds the current CPU number; this is the only CPU
925 * which processes crash_kexec routines.
926 */
927void __noclone __crash_kexec(struct pt_regs *regs)
928{
929 /* Take the kexec_mutex here to prevent sys_kexec_load
930 * running on one cpu from replacing the crash kernel
931 * we are using after a panic on a different cpu.
932 *
933 * If the crash kernel was not located in a fixed area
934 * of memory the xchg(&kexec_crash_image) would be
935 * sufficient. But since I reuse the memory...
936 */
937 if (mutex_trylock(&kexec_mutex)) {
938 if (kexec_crash_image) {
939 struct pt_regs fixed_regs;
940
941 crash_setup_regs(&fixed_regs, regs);
942 crash_save_vmcoreinfo();
943 machine_crash_shutdown(&fixed_regs);
944 machine_kexec(kexec_crash_image);
945 }
946 mutex_unlock(&kexec_mutex);
947 }
948}
949STACK_FRAME_NON_STANDARD(__crash_kexec);
950
951void crash_kexec(struct pt_regs *regs)
952{
953 int old_cpu, this_cpu;
954
955 /*
956 * Only one CPU is allowed to execute the crash_kexec() code as with
957 * panic(). Otherwise parallel calls of panic() and crash_kexec()
958 * may stop each other. To exclude them, we use panic_cpu here too.
959 */
960 this_cpu = raw_smp_processor_id();
961 old_cpu = atomic_cmpxchg(&panic_cpu, PANIC_CPU_INVALID, this_cpu);
962 if (old_cpu == PANIC_CPU_INVALID) {
963 /* This is the 1st CPU which comes here, so go ahead. */
964 printk_safe_flush_on_panic();
965 __crash_kexec(regs);
966
967 /*
968 * Reset panic_cpu to allow another panic()/crash_kexec()
969 * call.
970 */
971 atomic_set(&panic_cpu, PANIC_CPU_INVALID);
972 }
973}
974
975size_t crash_get_memory_size(void)
976{
977 size_t size = 0;
978
979 mutex_lock(&kexec_mutex);
980 if (crashk_res.end != crashk_res.start)
981 size = resource_size(&crashk_res);
982 mutex_unlock(&kexec_mutex);
983 return size;
984}
985
986void __weak crash_free_reserved_phys_range(unsigned long begin,
987 unsigned long end)
988{
989 unsigned long addr;
990
991 for (addr = begin; addr < end; addr += PAGE_SIZE)
992 free_reserved_page(boot_pfn_to_page(addr >> PAGE_SHIFT));
993}
994
995int crash_shrink_memory(unsigned long new_size)
996{
997 int ret = 0;
998 unsigned long start, end;
999 unsigned long old_size;
1000 struct resource *ram_res;
1001
1002 mutex_lock(&kexec_mutex);
1003
1004 if (kexec_crash_image) {
1005 ret = -ENOENT;
1006 goto unlock;
1007 }
1008 start = crashk_res.start;
1009 end = crashk_res.end;
1010 old_size = (end == 0) ? 0 : end - start + 1;
1011 if (new_size >= old_size) {
1012 ret = (new_size == old_size) ? 0 : -EINVAL;
1013 goto unlock;
1014 }
1015
1016 ram_res = kzalloc(sizeof(*ram_res), GFP_KERNEL);
1017 if (!ram_res) {
1018 ret = -ENOMEM;
1019 goto unlock;
1020 }
1021
1022 start = roundup(start, KEXEC_CRASH_MEM_ALIGN);
1023 end = roundup(start + new_size, KEXEC_CRASH_MEM_ALIGN);
1024
1025 crash_free_reserved_phys_range(end, crashk_res.end);
1026
1027 if ((start == end) && (crashk_res.parent != NULL))
1028 release_resource(&crashk_res);
1029
1030 ram_res->start = end;
1031 ram_res->end = crashk_res.end;
1032 ram_res->flags = IORESOURCE_BUSY | IORESOURCE_SYSTEM_RAM;
1033 ram_res->name = "System RAM";
1034
1035 crashk_res.end = end - 1;
1036
1037 insert_resource(&iomem_resource, ram_res);
1038
1039unlock:
1040 mutex_unlock(&kexec_mutex);
1041 return ret;
1042}
1043
1044void crash_save_cpu(struct pt_regs *regs, int cpu)
1045{
1046 struct elf_prstatus prstatus;
1047 u32 *buf;
1048
1049 if ((cpu < 0) || (cpu >= nr_cpu_ids))
1050 return;
1051
1052 /* Using ELF notes here is opportunistic.
1053 * I need a well defined structure format
1054 * for the data I pass, and I need tags
1055 * on the data to indicate what information I have
1056 * squirrelled away. ELF notes happen to provide
1057 * all of that, so there is no need to invent something new.
1058 */
1059 buf = (u32 *)per_cpu_ptr(crash_notes, cpu);
1060 if (!buf)
1061 return;
1062 memset(&prstatus, 0, sizeof(prstatus));
1063 prstatus.pr_pid = current->pid;
1064 elf_core_copy_kernel_regs(&prstatus.pr_reg, regs);
1065 buf = append_elf_note(buf, KEXEC_CORE_NOTE_NAME, NT_PRSTATUS,
1066 &prstatus, sizeof(prstatus));
1067 final_note(buf);
1068}
1069
1070static int __init crash_notes_memory_init(void)
1071{
1072 /* Allocate memory for saving cpu registers. */
1073 size_t size, align;
1074
1075 /*
1076 * crash_notes could be allocated across 2 vmalloc pages when percpu
1077 * is vmalloc based . vmalloc doesn't guarantee 2 continuous vmalloc
1078 * pages are also on 2 continuous physical pages. In this case the
1079 * 2nd part of crash_notes in 2nd page could be lost since only the
1080 * starting address and size of crash_notes are exported through sysfs.
1081 * Here round up the size of crash_notes to the nearest power of two
1082 * and pass it to __alloc_percpu as align value. This can make sure
1083 * crash_notes is allocated inside one physical page.
1084 */
1085 size = sizeof(note_buf_t);
1086 align = min(roundup_pow_of_two(sizeof(note_buf_t)), PAGE_SIZE);
1087
1088 /*
1089 * Break compile if size is bigger than PAGE_SIZE since crash_notes
1090 * definitely will be in 2 pages with that.
1091 */
1092 BUILD_BUG_ON(size > PAGE_SIZE);
1093
1094 crash_notes = __alloc_percpu(size, align);
1095 if (!crash_notes) {
1096 pr_warn("Memory allocation for saving cpu register states failed\n");
1097 return -ENOMEM;
1098 }
1099 return 0;
1100}
1101subsys_initcall(crash_notes_memory_init);
1102
1103
1104/*
1105 * Move into place and start executing a preloaded standalone
1106 * executable. If nothing was preloaded return an error.
1107 */
1108int kernel_kexec(void)
1109{
1110 int error = 0;
1111
1112 if (!mutex_trylock(&kexec_mutex))
1113 return -EBUSY;
1114 if (!kexec_image) {
1115 error = -EINVAL;
1116 goto Unlock;
1117 }
1118
1119#ifdef CONFIG_KEXEC_JUMP
1120 if (kexec_image->preserve_context) {
1121 lock_system_sleep();
1122 pm_prepare_console();
1123 error = freeze_processes();
1124 if (error) {
1125 error = -EBUSY;
1126 goto Restore_console;
1127 }
1128 suspend_console();
1129 error = dpm_suspend_start(PMSG_FREEZE);
1130 if (error)
1131 goto Resume_console;
1132 /* At this point, dpm_suspend_start() has been called,
1133 * but *not* dpm_suspend_end(). We *must* call
1134 * dpm_suspend_end() now. Otherwise, drivers for
1135 * some devices (e.g. interrupt controllers) become
1136 * desynchronized with the actual state of the
1137 * hardware at resume time, and evil weirdness ensues.
1138 */
1139 error = dpm_suspend_end(PMSG_FREEZE);
1140 if (error)
1141 goto Resume_devices;
1142 error = disable_nonboot_cpus();
1143 if (error)
1144 goto Enable_cpus;
1145 local_irq_disable();
1146 error = syscore_suspend();
1147 if (error)
1148 goto Enable_irqs;
1149 } else
1150#endif
1151 {
1152 kexec_in_progress = true;
1153 kernel_restart_prepare(NULL);
1154 migrate_to_reboot_cpu();
1155
1156 /*
1157 * migrate_to_reboot_cpu() disables CPU hotplug assuming that
1158 * no further code needs to use CPU hotplug (which is true in
1159 * the reboot case). However, the kexec path depends on using
1160 * CPU hotplug again; so re-enable it here.
1161 */
1162 cpu_hotplug_enable();
1163 pr_emerg("Starting new kernel\n");
1164 machine_shutdown();
1165 }
1166
1167 machine_kexec(kexec_image);
1168
1169#ifdef CONFIG_KEXEC_JUMP
1170 if (kexec_image->preserve_context) {
1171 syscore_resume();
1172 Enable_irqs:
1173 local_irq_enable();
1174 Enable_cpus:
1175 enable_nonboot_cpus();
1176 dpm_resume_start(PMSG_RESTORE);
1177 Resume_devices:
1178 dpm_resume_end(PMSG_RESTORE);
1179 Resume_console:
1180 resume_console();
1181 thaw_processes();
1182 Restore_console:
1183 pm_restore_console();
1184 unlock_system_sleep();
1185 }
1186#endif
1187
1188 Unlock:
1189 mutex_unlock(&kexec_mutex);
1190 return error;
1191}
1192
1193/*
1194 * Protection mechanism for crashkernel reserved memory after
1195 * the kdump kernel is loaded.
1196 *
1197 * Provide an empty default implementation here -- architecture
1198 * code may override this
1199 */
1200void __weak arch_kexec_protect_crashkres(void)
1201{}
1202
1203void __weak arch_kexec_unprotect_crashkres(void)
1204{}
1// SPDX-License-Identifier: GPL-2.0-only
2/*
3 * kexec.c - kexec system call core code.
4 * Copyright (C) 2002-2004 Eric Biederman <ebiederm@xmission.com>
5 */
6
7#define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
8
9#include <linux/btf.h>
10#include <linux/capability.h>
11#include <linux/mm.h>
12#include <linux/file.h>
13#include <linux/slab.h>
14#include <linux/fs.h>
15#include <linux/kexec.h>
16#include <linux/mutex.h>
17#include <linux/list.h>
18#include <linux/highmem.h>
19#include <linux/syscalls.h>
20#include <linux/reboot.h>
21#include <linux/ioport.h>
22#include <linux/hardirq.h>
23#include <linux/elf.h>
24#include <linux/elfcore.h>
25#include <linux/utsname.h>
26#include <linux/numa.h>
27#include <linux/suspend.h>
28#include <linux/device.h>
29#include <linux/freezer.h>
30#include <linux/panic_notifier.h>
31#include <linux/pm.h>
32#include <linux/cpu.h>
33#include <linux/uaccess.h>
34#include <linux/io.h>
35#include <linux/console.h>
36#include <linux/vmalloc.h>
37#include <linux/swap.h>
38#include <linux/syscore_ops.h>
39#include <linux/compiler.h>
40#include <linux/hugetlb.h>
41#include <linux/objtool.h>
42#include <linux/kmsg_dump.h>
43
44#include <asm/page.h>
45#include <asm/sections.h>
46
47#include <crypto/hash.h>
48#include "kexec_internal.h"
49
50atomic_t __kexec_lock = ATOMIC_INIT(0);
51
52/* Flag to indicate we are going to kexec a new kernel */
53bool kexec_in_progress = false;
54
55bool kexec_file_dbg_print;
56
57/*
58 * When kexec transitions to the new kernel there is a one-to-one
59 * mapping between physical and virtual addresses. On processors
60 * where you can disable the MMU this is trivial, and easy. For
61 * others it is still a simple predictable page table to setup.
62 *
63 * In that environment kexec copies the new kernel to its final
64 * resting place. This means I can only support memory whose
65 * physical address can fit in an unsigned long. In particular
66 * addresses where (pfn << PAGE_SHIFT) > ULONG_MAX cannot be handled.
67 * If the assembly stub has more restrictive requirements
68 * KEXEC_SOURCE_MEMORY_LIMIT and KEXEC_DEST_MEMORY_LIMIT can be
69 * defined more restrictively in <asm/kexec.h>.
70 *
71 * The code for the transition from the current kernel to the
72 * new kernel is placed in the control_code_buffer, whose size
73 * is given by KEXEC_CONTROL_PAGE_SIZE. In the best case only a single
74 * page of memory is necessary, but some architectures require more.
75 * Because this memory must be identity mapped in the transition from
76 * virtual to physical addresses it must live in the range
77 * 0 - TASK_SIZE, as only the user space mappings are arbitrarily
78 * modifiable.
79 *
80 * The assembly stub in the control code buffer is passed a linked list
81 * of descriptor pages detailing the source pages of the new kernel,
82 * and the destination addresses of those source pages. As this data
83 * structure is not used in the context of the current OS, it must
84 * be self-contained.
85 *
86 * The code has been made to work with highmem pages and will use a
87 * destination page in its final resting place (if it happens
88 * to allocate it). The end product of this is that most of the
89 * physical address space, and most of RAM can be used.
90 *
91 * Future directions include:
92 * - allocating a page table with the control code buffer identity
93 * mapped, to simplify machine_kexec and make kexec_on_panic more
94 * reliable.
95 */
96
97/*
98 * KIMAGE_NO_DEST is an impossible destination address..., for
99 * allocating pages whose destination address we do not care about.
100 */
101#define KIMAGE_NO_DEST (-1UL)
102#define PAGE_COUNT(x) (((x) + PAGE_SIZE - 1) >> PAGE_SHIFT)
103
104static struct page *kimage_alloc_page(struct kimage *image,
105 gfp_t gfp_mask,
106 unsigned long dest);
107
108int sanity_check_segment_list(struct kimage *image)
109{
110 int i;
111 unsigned long nr_segments = image->nr_segments;
112 unsigned long total_pages = 0;
113 unsigned long nr_pages = totalram_pages();
114
115 /*
116 * Verify we have good destination addresses. The caller is
117 * responsible for making certain we don't attempt to load
118 * the new image into invalid or reserved areas of RAM. This
119 * just verifies it is an address we can use.
120 *
121 * Since the kernel does everything in page size chunks ensure
122 * the destination addresses are page aligned. Too many
123 * special cases crop of when we don't do this. The most
124 * insidious is getting overlapping destination addresses
125 * simply because addresses are changed to page size
126 * granularity.
127 */
128 for (i = 0; i < nr_segments; i++) {
129 unsigned long mstart, mend;
130
131 mstart = image->segment[i].mem;
132 mend = mstart + image->segment[i].memsz;
133 if (mstart > mend)
134 return -EADDRNOTAVAIL;
135 if ((mstart & ~PAGE_MASK) || (mend & ~PAGE_MASK))
136 return -EADDRNOTAVAIL;
137 if (mend >= KEXEC_DESTINATION_MEMORY_LIMIT)
138 return -EADDRNOTAVAIL;
139 }
140
141 /* Verify our destination addresses do not overlap.
142 * If we alloed overlapping destination addresses
143 * through very weird things can happen with no
144 * easy explanation as one segment stops on another.
145 */
146 for (i = 0; i < nr_segments; i++) {
147 unsigned long mstart, mend;
148 unsigned long j;
149
150 mstart = image->segment[i].mem;
151 mend = mstart + image->segment[i].memsz;
152 for (j = 0; j < i; j++) {
153 unsigned long pstart, pend;
154
155 pstart = image->segment[j].mem;
156 pend = pstart + image->segment[j].memsz;
157 /* Do the segments overlap ? */
158 if ((mend > pstart) && (mstart < pend))
159 return -EINVAL;
160 }
161 }
162
163 /* Ensure our buffer sizes are strictly less than
164 * our memory sizes. This should always be the case,
165 * and it is easier to check up front than to be surprised
166 * later on.
167 */
168 for (i = 0; i < nr_segments; i++) {
169 if (image->segment[i].bufsz > image->segment[i].memsz)
170 return -EINVAL;
171 }
172
173 /*
174 * Verify that no more than half of memory will be consumed. If the
175 * request from userspace is too large, a large amount of time will be
176 * wasted allocating pages, which can cause a soft lockup.
177 */
178 for (i = 0; i < nr_segments; i++) {
179 if (PAGE_COUNT(image->segment[i].memsz) > nr_pages / 2)
180 return -EINVAL;
181
182 total_pages += PAGE_COUNT(image->segment[i].memsz);
183 }
184
185 if (total_pages > nr_pages / 2)
186 return -EINVAL;
187
188#ifdef CONFIG_CRASH_DUMP
189 /*
190 * Verify we have good destination addresses. Normally
191 * the caller is responsible for making certain we don't
192 * attempt to load the new image into invalid or reserved
193 * areas of RAM. But crash kernels are preloaded into a
194 * reserved area of ram. We must ensure the addresses
195 * are in the reserved area otherwise preloading the
196 * kernel could corrupt things.
197 */
198
199 if (image->type == KEXEC_TYPE_CRASH) {
200 for (i = 0; i < nr_segments; i++) {
201 unsigned long mstart, mend;
202
203 mstart = image->segment[i].mem;
204 mend = mstart + image->segment[i].memsz - 1;
205 /* Ensure we are within the crash kernel limits */
206 if ((mstart < phys_to_boot_phys(crashk_res.start)) ||
207 (mend > phys_to_boot_phys(crashk_res.end)))
208 return -EADDRNOTAVAIL;
209 }
210 }
211#endif
212
213 return 0;
214}
215
216struct kimage *do_kimage_alloc_init(void)
217{
218 struct kimage *image;
219
220 /* Allocate a controlling structure */
221 image = kzalloc(sizeof(*image), GFP_KERNEL);
222 if (!image)
223 return NULL;
224
225 image->head = 0;
226 image->entry = &image->head;
227 image->last_entry = &image->head;
228 image->control_page = ~0; /* By default this does not apply */
229 image->type = KEXEC_TYPE_DEFAULT;
230
231 /* Initialize the list of control pages */
232 INIT_LIST_HEAD(&image->control_pages);
233
234 /* Initialize the list of destination pages */
235 INIT_LIST_HEAD(&image->dest_pages);
236
237 /* Initialize the list of unusable pages */
238 INIT_LIST_HEAD(&image->unusable_pages);
239
240#ifdef CONFIG_CRASH_HOTPLUG
241 image->hp_action = KEXEC_CRASH_HP_NONE;
242 image->elfcorehdr_index = -1;
243 image->elfcorehdr_updated = false;
244#endif
245
246 return image;
247}
248
249int kimage_is_destination_range(struct kimage *image,
250 unsigned long start,
251 unsigned long end)
252{
253 unsigned long i;
254
255 for (i = 0; i < image->nr_segments; i++) {
256 unsigned long mstart, mend;
257
258 mstart = image->segment[i].mem;
259 mend = mstart + image->segment[i].memsz - 1;
260 if ((end >= mstart) && (start <= mend))
261 return 1;
262 }
263
264 return 0;
265}
266
267static struct page *kimage_alloc_pages(gfp_t gfp_mask, unsigned int order)
268{
269 struct page *pages;
270
271 if (fatal_signal_pending(current))
272 return NULL;
273 pages = alloc_pages(gfp_mask & ~__GFP_ZERO, order);
274 if (pages) {
275 unsigned int count, i;
276
277 pages->mapping = NULL;
278 set_page_private(pages, order);
279 count = 1 << order;
280 for (i = 0; i < count; i++)
281 SetPageReserved(pages + i);
282
283 arch_kexec_post_alloc_pages(page_address(pages), count,
284 gfp_mask);
285
286 if (gfp_mask & __GFP_ZERO)
287 for (i = 0; i < count; i++)
288 clear_highpage(pages + i);
289 }
290
291 return pages;
292}
293
294static void kimage_free_pages(struct page *page)
295{
296 unsigned int order, count, i;
297
298 order = page_private(page);
299 count = 1 << order;
300
301 arch_kexec_pre_free_pages(page_address(page), count);
302
303 for (i = 0; i < count; i++)
304 ClearPageReserved(page + i);
305 __free_pages(page, order);
306}
307
308void kimage_free_page_list(struct list_head *list)
309{
310 struct page *page, *next;
311
312 list_for_each_entry_safe(page, next, list, lru) {
313 list_del(&page->lru);
314 kimage_free_pages(page);
315 }
316}
317
318static struct page *kimage_alloc_normal_control_pages(struct kimage *image,
319 unsigned int order)
320{
321 /* Control pages are special, they are the intermediaries
322 * that are needed while we copy the rest of the pages
323 * to their final resting place. As such they must
324 * not conflict with either the destination addresses
325 * or memory the kernel is already using.
326 *
327 * The only case where we really need more than one of
328 * these are for architectures where we cannot disable
329 * the MMU and must instead generate an identity mapped
330 * page table for all of the memory.
331 *
332 * At worst this runs in O(N) of the image size.
333 */
334 struct list_head extra_pages;
335 struct page *pages;
336 unsigned int count;
337
338 count = 1 << order;
339 INIT_LIST_HEAD(&extra_pages);
340
341 /* Loop while I can allocate a page and the page allocated
342 * is a destination page.
343 */
344 do {
345 unsigned long pfn, epfn, addr, eaddr;
346
347 pages = kimage_alloc_pages(KEXEC_CONTROL_MEMORY_GFP, order);
348 if (!pages)
349 break;
350 pfn = page_to_boot_pfn(pages);
351 epfn = pfn + count;
352 addr = pfn << PAGE_SHIFT;
353 eaddr = (epfn << PAGE_SHIFT) - 1;
354 if ((epfn >= (KEXEC_CONTROL_MEMORY_LIMIT >> PAGE_SHIFT)) ||
355 kimage_is_destination_range(image, addr, eaddr)) {
356 list_add(&pages->lru, &extra_pages);
357 pages = NULL;
358 }
359 } while (!pages);
360
361 if (pages) {
362 /* Remember the allocated page... */
363 list_add(&pages->lru, &image->control_pages);
364
365 /* Because the page is already in it's destination
366 * location we will never allocate another page at
367 * that address. Therefore kimage_alloc_pages
368 * will not return it (again) and we don't need
369 * to give it an entry in image->segment[].
370 */
371 }
372 /* Deal with the destination pages I have inadvertently allocated.
373 *
374 * Ideally I would convert multi-page allocations into single
375 * page allocations, and add everything to image->dest_pages.
376 *
377 * For now it is simpler to just free the pages.
378 */
379 kimage_free_page_list(&extra_pages);
380
381 return pages;
382}
383
384#ifdef CONFIG_CRASH_DUMP
385static struct page *kimage_alloc_crash_control_pages(struct kimage *image,
386 unsigned int order)
387{
388 /* Control pages are special, they are the intermediaries
389 * that are needed while we copy the rest of the pages
390 * to their final resting place. As such they must
391 * not conflict with either the destination addresses
392 * or memory the kernel is already using.
393 *
394 * Control pages are also the only pags we must allocate
395 * when loading a crash kernel. All of the other pages
396 * are specified by the segments and we just memcpy
397 * into them directly.
398 *
399 * The only case where we really need more than one of
400 * these are for architectures where we cannot disable
401 * the MMU and must instead generate an identity mapped
402 * page table for all of the memory.
403 *
404 * Given the low demand this implements a very simple
405 * allocator that finds the first hole of the appropriate
406 * size in the reserved memory region, and allocates all
407 * of the memory up to and including the hole.
408 */
409 unsigned long hole_start, hole_end, size;
410 struct page *pages;
411
412 pages = NULL;
413 size = (1 << order) << PAGE_SHIFT;
414 hole_start = ALIGN(image->control_page, size);
415 hole_end = hole_start + size - 1;
416 while (hole_end <= crashk_res.end) {
417 unsigned long i;
418
419 cond_resched();
420
421 if (hole_end > KEXEC_CRASH_CONTROL_MEMORY_LIMIT)
422 break;
423 /* See if I overlap any of the segments */
424 for (i = 0; i < image->nr_segments; i++) {
425 unsigned long mstart, mend;
426
427 mstart = image->segment[i].mem;
428 mend = mstart + image->segment[i].memsz - 1;
429 if ((hole_end >= mstart) && (hole_start <= mend)) {
430 /* Advance the hole to the end of the segment */
431 hole_start = ALIGN(mend, size);
432 hole_end = hole_start + size - 1;
433 break;
434 }
435 }
436 /* If I don't overlap any segments I have found my hole! */
437 if (i == image->nr_segments) {
438 pages = pfn_to_page(hole_start >> PAGE_SHIFT);
439 image->control_page = hole_end + 1;
440 break;
441 }
442 }
443
444 /* Ensure that these pages are decrypted if SME is enabled. */
445 if (pages)
446 arch_kexec_post_alloc_pages(page_address(pages), 1 << order, 0);
447
448 return pages;
449}
450#endif
451
452
453struct page *kimage_alloc_control_pages(struct kimage *image,
454 unsigned int order)
455{
456 struct page *pages = NULL;
457
458 switch (image->type) {
459 case KEXEC_TYPE_DEFAULT:
460 pages = kimage_alloc_normal_control_pages(image, order);
461 break;
462#ifdef CONFIG_CRASH_DUMP
463 case KEXEC_TYPE_CRASH:
464 pages = kimage_alloc_crash_control_pages(image, order);
465 break;
466#endif
467 }
468
469 return pages;
470}
471
472static int kimage_add_entry(struct kimage *image, kimage_entry_t entry)
473{
474 if (*image->entry != 0)
475 image->entry++;
476
477 if (image->entry == image->last_entry) {
478 kimage_entry_t *ind_page;
479 struct page *page;
480
481 page = kimage_alloc_page(image, GFP_KERNEL, KIMAGE_NO_DEST);
482 if (!page)
483 return -ENOMEM;
484
485 ind_page = page_address(page);
486 *image->entry = virt_to_boot_phys(ind_page) | IND_INDIRECTION;
487 image->entry = ind_page;
488 image->last_entry = ind_page +
489 ((PAGE_SIZE/sizeof(kimage_entry_t)) - 1);
490 }
491 *image->entry = entry;
492 image->entry++;
493 *image->entry = 0;
494
495 return 0;
496}
497
498static int kimage_set_destination(struct kimage *image,
499 unsigned long destination)
500{
501 destination &= PAGE_MASK;
502
503 return kimage_add_entry(image, destination | IND_DESTINATION);
504}
505
506
507static int kimage_add_page(struct kimage *image, unsigned long page)
508{
509 page &= PAGE_MASK;
510
511 return kimage_add_entry(image, page | IND_SOURCE);
512}
513
514
515static void kimage_free_extra_pages(struct kimage *image)
516{
517 /* Walk through and free any extra destination pages I may have */
518 kimage_free_page_list(&image->dest_pages);
519
520 /* Walk through and free any unusable pages I have cached */
521 kimage_free_page_list(&image->unusable_pages);
522
523}
524
525void kimage_terminate(struct kimage *image)
526{
527 if (*image->entry != 0)
528 image->entry++;
529
530 *image->entry = IND_DONE;
531}
532
533#define for_each_kimage_entry(image, ptr, entry) \
534 for (ptr = &image->head; (entry = *ptr) && !(entry & IND_DONE); \
535 ptr = (entry & IND_INDIRECTION) ? \
536 boot_phys_to_virt((entry & PAGE_MASK)) : ptr + 1)
537
538static void kimage_free_entry(kimage_entry_t entry)
539{
540 struct page *page;
541
542 page = boot_pfn_to_page(entry >> PAGE_SHIFT);
543 kimage_free_pages(page);
544}
545
546void kimage_free(struct kimage *image)
547{
548 kimage_entry_t *ptr, entry;
549 kimage_entry_t ind = 0;
550
551 if (!image)
552 return;
553
554#ifdef CONFIG_CRASH_DUMP
555 if (image->vmcoreinfo_data_copy) {
556 crash_update_vmcoreinfo_safecopy(NULL);
557 vunmap(image->vmcoreinfo_data_copy);
558 }
559#endif
560
561 kimage_free_extra_pages(image);
562 for_each_kimage_entry(image, ptr, entry) {
563 if (entry & IND_INDIRECTION) {
564 /* Free the previous indirection page */
565 if (ind & IND_INDIRECTION)
566 kimage_free_entry(ind);
567 /* Save this indirection page until we are
568 * done with it.
569 */
570 ind = entry;
571 } else if (entry & IND_SOURCE)
572 kimage_free_entry(entry);
573 }
574 /* Free the final indirection page */
575 if (ind & IND_INDIRECTION)
576 kimage_free_entry(ind);
577
578 /* Handle any machine specific cleanup */
579 machine_kexec_cleanup(image);
580
581 /* Free the kexec control pages... */
582 kimage_free_page_list(&image->control_pages);
583
584 /*
585 * Free up any temporary buffers allocated. This might hit if
586 * error occurred much later after buffer allocation.
587 */
588 if (image->file_mode)
589 kimage_file_post_load_cleanup(image);
590
591 kfree(image);
592}
593
594static kimage_entry_t *kimage_dst_used(struct kimage *image,
595 unsigned long page)
596{
597 kimage_entry_t *ptr, entry;
598 unsigned long destination = 0;
599
600 for_each_kimage_entry(image, ptr, entry) {
601 if (entry & IND_DESTINATION)
602 destination = entry & PAGE_MASK;
603 else if (entry & IND_SOURCE) {
604 if (page == destination)
605 return ptr;
606 destination += PAGE_SIZE;
607 }
608 }
609
610 return NULL;
611}
612
613static struct page *kimage_alloc_page(struct kimage *image,
614 gfp_t gfp_mask,
615 unsigned long destination)
616{
617 /*
618 * Here we implement safeguards to ensure that a source page
619 * is not copied to its destination page before the data on
620 * the destination page is no longer useful.
621 *
622 * To do this we maintain the invariant that a source page is
623 * either its own destination page, or it is not a
624 * destination page at all.
625 *
626 * That is slightly stronger than required, but the proof
627 * that no problems will not occur is trivial, and the
628 * implementation is simply to verify.
629 *
630 * When allocating all pages normally this algorithm will run
631 * in O(N) time, but in the worst case it will run in O(N^2)
632 * time. If the runtime is a problem the data structures can
633 * be fixed.
634 */
635 struct page *page;
636 unsigned long addr;
637
638 /*
639 * Walk through the list of destination pages, and see if I
640 * have a match.
641 */
642 list_for_each_entry(page, &image->dest_pages, lru) {
643 addr = page_to_boot_pfn(page) << PAGE_SHIFT;
644 if (addr == destination) {
645 list_del(&page->lru);
646 return page;
647 }
648 }
649 page = NULL;
650 while (1) {
651 kimage_entry_t *old;
652
653 /* Allocate a page, if we run out of memory give up */
654 page = kimage_alloc_pages(gfp_mask, 0);
655 if (!page)
656 return NULL;
657 /* If the page cannot be used file it away */
658 if (page_to_boot_pfn(page) >
659 (KEXEC_SOURCE_MEMORY_LIMIT >> PAGE_SHIFT)) {
660 list_add(&page->lru, &image->unusable_pages);
661 continue;
662 }
663 addr = page_to_boot_pfn(page) << PAGE_SHIFT;
664
665 /* If it is the destination page we want use it */
666 if (addr == destination)
667 break;
668
669 /* If the page is not a destination page use it */
670 if (!kimage_is_destination_range(image, addr,
671 addr + PAGE_SIZE - 1))
672 break;
673
674 /*
675 * I know that the page is someones destination page.
676 * See if there is already a source page for this
677 * destination page. And if so swap the source pages.
678 */
679 old = kimage_dst_used(image, addr);
680 if (old) {
681 /* If so move it */
682 unsigned long old_addr;
683 struct page *old_page;
684
685 old_addr = *old & PAGE_MASK;
686 old_page = boot_pfn_to_page(old_addr >> PAGE_SHIFT);
687 copy_highpage(page, old_page);
688 *old = addr | (*old & ~PAGE_MASK);
689
690 /* The old page I have found cannot be a
691 * destination page, so return it if it's
692 * gfp_flags honor the ones passed in.
693 */
694 if (!(gfp_mask & __GFP_HIGHMEM) &&
695 PageHighMem(old_page)) {
696 kimage_free_pages(old_page);
697 continue;
698 }
699 page = old_page;
700 break;
701 }
702 /* Place the page on the destination list, to be used later */
703 list_add(&page->lru, &image->dest_pages);
704 }
705
706 return page;
707}
708
709static int kimage_load_normal_segment(struct kimage *image,
710 struct kexec_segment *segment)
711{
712 unsigned long maddr;
713 size_t ubytes, mbytes;
714 int result;
715 unsigned char __user *buf = NULL;
716 unsigned char *kbuf = NULL;
717
718 if (image->file_mode)
719 kbuf = segment->kbuf;
720 else
721 buf = segment->buf;
722 ubytes = segment->bufsz;
723 mbytes = segment->memsz;
724 maddr = segment->mem;
725
726 result = kimage_set_destination(image, maddr);
727 if (result < 0)
728 goto out;
729
730 while (mbytes) {
731 struct page *page;
732 char *ptr;
733 size_t uchunk, mchunk;
734
735 page = kimage_alloc_page(image, GFP_HIGHUSER, maddr);
736 if (!page) {
737 result = -ENOMEM;
738 goto out;
739 }
740 result = kimage_add_page(image, page_to_boot_pfn(page)
741 << PAGE_SHIFT);
742 if (result < 0)
743 goto out;
744
745 ptr = kmap_local_page(page);
746 /* Start with a clear page */
747 clear_page(ptr);
748 ptr += maddr & ~PAGE_MASK;
749 mchunk = min_t(size_t, mbytes,
750 PAGE_SIZE - (maddr & ~PAGE_MASK));
751 uchunk = min(ubytes, mchunk);
752
753 if (uchunk) {
754 /* For file based kexec, source pages are in kernel memory */
755 if (image->file_mode)
756 memcpy(ptr, kbuf, uchunk);
757 else
758 result = copy_from_user(ptr, buf, uchunk);
759 ubytes -= uchunk;
760 if (image->file_mode)
761 kbuf += uchunk;
762 else
763 buf += uchunk;
764 }
765 kunmap_local(ptr);
766 if (result) {
767 result = -EFAULT;
768 goto out;
769 }
770 maddr += mchunk;
771 mbytes -= mchunk;
772
773 cond_resched();
774 }
775out:
776 return result;
777}
778
779#ifdef CONFIG_CRASH_DUMP
780static int kimage_load_crash_segment(struct kimage *image,
781 struct kexec_segment *segment)
782{
783 /* For crash dumps kernels we simply copy the data from
784 * user space to it's destination.
785 * We do things a page at a time for the sake of kmap.
786 */
787 unsigned long maddr;
788 size_t ubytes, mbytes;
789 int result;
790 unsigned char __user *buf = NULL;
791 unsigned char *kbuf = NULL;
792
793 result = 0;
794 if (image->file_mode)
795 kbuf = segment->kbuf;
796 else
797 buf = segment->buf;
798 ubytes = segment->bufsz;
799 mbytes = segment->memsz;
800 maddr = segment->mem;
801 while (mbytes) {
802 struct page *page;
803 char *ptr;
804 size_t uchunk, mchunk;
805
806 page = boot_pfn_to_page(maddr >> PAGE_SHIFT);
807 if (!page) {
808 result = -ENOMEM;
809 goto out;
810 }
811 arch_kexec_post_alloc_pages(page_address(page), 1, 0);
812 ptr = kmap_local_page(page);
813 ptr += maddr & ~PAGE_MASK;
814 mchunk = min_t(size_t, mbytes,
815 PAGE_SIZE - (maddr & ~PAGE_MASK));
816 uchunk = min(ubytes, mchunk);
817 if (mchunk > uchunk) {
818 /* Zero the trailing part of the page */
819 memset(ptr + uchunk, 0, mchunk - uchunk);
820 }
821
822 if (uchunk) {
823 /* For file based kexec, source pages are in kernel memory */
824 if (image->file_mode)
825 memcpy(ptr, kbuf, uchunk);
826 else
827 result = copy_from_user(ptr, buf, uchunk);
828 ubytes -= uchunk;
829 if (image->file_mode)
830 kbuf += uchunk;
831 else
832 buf += uchunk;
833 }
834 kexec_flush_icache_page(page);
835 kunmap_local(ptr);
836 arch_kexec_pre_free_pages(page_address(page), 1);
837 if (result) {
838 result = -EFAULT;
839 goto out;
840 }
841 maddr += mchunk;
842 mbytes -= mchunk;
843
844 cond_resched();
845 }
846out:
847 return result;
848}
849#endif
850
851int kimage_load_segment(struct kimage *image,
852 struct kexec_segment *segment)
853{
854 int result = -ENOMEM;
855
856 switch (image->type) {
857 case KEXEC_TYPE_DEFAULT:
858 result = kimage_load_normal_segment(image, segment);
859 break;
860#ifdef CONFIG_CRASH_DUMP
861 case KEXEC_TYPE_CRASH:
862 result = kimage_load_crash_segment(image, segment);
863 break;
864#endif
865 }
866
867 return result;
868}
869
870struct kexec_load_limit {
871 /* Mutex protects the limit count. */
872 struct mutex mutex;
873 int limit;
874};
875
876static struct kexec_load_limit load_limit_reboot = {
877 .mutex = __MUTEX_INITIALIZER(load_limit_reboot.mutex),
878 .limit = -1,
879};
880
881static struct kexec_load_limit load_limit_panic = {
882 .mutex = __MUTEX_INITIALIZER(load_limit_panic.mutex),
883 .limit = -1,
884};
885
886struct kimage *kexec_image;
887struct kimage *kexec_crash_image;
888static int kexec_load_disabled;
889
890#ifdef CONFIG_SYSCTL
891static int kexec_limit_handler(struct ctl_table *table, int write,
892 void *buffer, size_t *lenp, loff_t *ppos)
893{
894 struct kexec_load_limit *limit = table->data;
895 int val;
896 struct ctl_table tmp = {
897 .data = &val,
898 .maxlen = sizeof(val),
899 .mode = table->mode,
900 };
901 int ret;
902
903 if (write) {
904 ret = proc_dointvec(&tmp, write, buffer, lenp, ppos);
905 if (ret)
906 return ret;
907
908 if (val < 0)
909 return -EINVAL;
910
911 mutex_lock(&limit->mutex);
912 if (limit->limit != -1 && val >= limit->limit)
913 ret = -EINVAL;
914 else
915 limit->limit = val;
916 mutex_unlock(&limit->mutex);
917
918 return ret;
919 }
920
921 mutex_lock(&limit->mutex);
922 val = limit->limit;
923 mutex_unlock(&limit->mutex);
924
925 return proc_dointvec(&tmp, write, buffer, lenp, ppos);
926}
927
928static struct ctl_table kexec_core_sysctls[] = {
929 {
930 .procname = "kexec_load_disabled",
931 .data = &kexec_load_disabled,
932 .maxlen = sizeof(int),
933 .mode = 0644,
934 /* only handle a transition from default "0" to "1" */
935 .proc_handler = proc_dointvec_minmax,
936 .extra1 = SYSCTL_ONE,
937 .extra2 = SYSCTL_ONE,
938 },
939 {
940 .procname = "kexec_load_limit_panic",
941 .data = &load_limit_panic,
942 .mode = 0644,
943 .proc_handler = kexec_limit_handler,
944 },
945 {
946 .procname = "kexec_load_limit_reboot",
947 .data = &load_limit_reboot,
948 .mode = 0644,
949 .proc_handler = kexec_limit_handler,
950 },
951 { }
952};
953
954static int __init kexec_core_sysctl_init(void)
955{
956 register_sysctl_init("kernel", kexec_core_sysctls);
957 return 0;
958}
959late_initcall(kexec_core_sysctl_init);
960#endif
961
962bool kexec_load_permitted(int kexec_image_type)
963{
964 struct kexec_load_limit *limit;
965
966 /*
967 * Only the superuser can use the kexec syscall and if it has not
968 * been disabled.
969 */
970 if (!capable(CAP_SYS_BOOT) || kexec_load_disabled)
971 return false;
972
973 /* Check limit counter and decrease it.*/
974 limit = (kexec_image_type == KEXEC_TYPE_CRASH) ?
975 &load_limit_panic : &load_limit_reboot;
976 mutex_lock(&limit->mutex);
977 if (!limit->limit) {
978 mutex_unlock(&limit->mutex);
979 return false;
980 }
981 if (limit->limit != -1)
982 limit->limit--;
983 mutex_unlock(&limit->mutex);
984
985 return true;
986}
987
988/*
989 * Move into place and start executing a preloaded standalone
990 * executable. If nothing was preloaded return an error.
991 */
992int kernel_kexec(void)
993{
994 int error = 0;
995
996 if (!kexec_trylock())
997 return -EBUSY;
998 if (!kexec_image) {
999 error = -EINVAL;
1000 goto Unlock;
1001 }
1002
1003#ifdef CONFIG_KEXEC_JUMP
1004 if (kexec_image->preserve_context) {
1005 pm_prepare_console();
1006 error = freeze_processes();
1007 if (error) {
1008 error = -EBUSY;
1009 goto Restore_console;
1010 }
1011 suspend_console();
1012 error = dpm_suspend_start(PMSG_FREEZE);
1013 if (error)
1014 goto Resume_console;
1015 /* At this point, dpm_suspend_start() has been called,
1016 * but *not* dpm_suspend_end(). We *must* call
1017 * dpm_suspend_end() now. Otherwise, drivers for
1018 * some devices (e.g. interrupt controllers) become
1019 * desynchronized with the actual state of the
1020 * hardware at resume time, and evil weirdness ensues.
1021 */
1022 error = dpm_suspend_end(PMSG_FREEZE);
1023 if (error)
1024 goto Resume_devices;
1025 error = suspend_disable_secondary_cpus();
1026 if (error)
1027 goto Enable_cpus;
1028 local_irq_disable();
1029 error = syscore_suspend();
1030 if (error)
1031 goto Enable_irqs;
1032 } else
1033#endif
1034 {
1035 kexec_in_progress = true;
1036 kernel_restart_prepare("kexec reboot");
1037 migrate_to_reboot_cpu();
1038 syscore_shutdown();
1039
1040 /*
1041 * migrate_to_reboot_cpu() disables CPU hotplug assuming that
1042 * no further code needs to use CPU hotplug (which is true in
1043 * the reboot case). However, the kexec path depends on using
1044 * CPU hotplug again; so re-enable it here.
1045 */
1046 cpu_hotplug_enable();
1047 pr_notice("Starting new kernel\n");
1048 machine_shutdown();
1049 }
1050
1051 kmsg_dump(KMSG_DUMP_SHUTDOWN);
1052 machine_kexec(kexec_image);
1053
1054#ifdef CONFIG_KEXEC_JUMP
1055 if (kexec_image->preserve_context) {
1056 syscore_resume();
1057 Enable_irqs:
1058 local_irq_enable();
1059 Enable_cpus:
1060 suspend_enable_secondary_cpus();
1061 dpm_resume_start(PMSG_RESTORE);
1062 Resume_devices:
1063 dpm_resume_end(PMSG_RESTORE);
1064 Resume_console:
1065 resume_console();
1066 thaw_processes();
1067 Restore_console:
1068 pm_restore_console();
1069 }
1070#endif
1071
1072 Unlock:
1073 kexec_unlock();
1074 return error;
1075}