1 // SPDX-License-Identifier: GPL-2.0-or-later
3 * Firmware Assisted dump: A robust mechanism to get reliable kernel crash
4 * dump with assistance from firmware. This approach does not use kexec,
5 * instead firmware assists in booting the kdump kernel while preserving
6 * memory contents. The most of the code implementation has been adapted
7 * from phyp assisted dump implementation written by Linas Vepstas and
10 * Copyright 2011 IBM Corporation
15 #define pr_fmt(fmt) "fadump: " fmt
17 #include <linux/string.h>
18 #include <linux/memblock.h>
19 #include <linux/delay.h>
20 #include <linux/seq_file.h>
21 #include <linux/crash_dump.h>
22 #include <linux/kobject.h>
23 #include <linux/sysfs.h>
24 #include <linux/slab.h>
25 #include <linux/cma.h>
26 #include <linux/hugetlb.h>
27 #include <linux/debugfs.h>
29 #include <linux/of_fdt.h>
32 #include <asm/fadump.h>
33 #include <asm/fadump-internal.h>
34 #include <asm/setup.h>
35 #include <asm/interrupt.h>
38 * The CPU who acquired the lock to trigger the fadump crash should
39 * wait for other CPUs to enter.
41 * The timeout is in milliseconds.
43 #define CRASH_TIMEOUT 500
45 static struct fw_dump fw_dump;
47 static void __init fadump_reserve_crash_area(u64 base);
49 #ifndef CONFIG_PRESERVE_FA_DUMP
51 static struct kobject *fadump_kobj;
53 static atomic_t cpus_in_fadump;
54 static DEFINE_MUTEX(fadump_mutex);
56 static struct fadump_mrange_info crash_mrange_info = { "crash", NULL, 0, 0, 0, false };
58 #define RESERVED_RNGS_SZ 16384 /* 16K - 128 entries */
59 #define RESERVED_RNGS_CNT (RESERVED_RNGS_SZ / \
60 sizeof(struct fadump_memory_range))
61 static struct fadump_memory_range rngs[RESERVED_RNGS_CNT];
62 static struct fadump_mrange_info
63 reserved_mrange_info = { "reserved", rngs, RESERVED_RNGS_SZ, 0, RESERVED_RNGS_CNT, true };
65 static void __init early_init_dt_scan_reserved_ranges(unsigned long node);
68 static struct cma *fadump_cma;
71 * fadump_cma_init() - Initialize CMA area from a fadump reserved memory
73 * This function initializes CMA area from fadump reserved memory.
74 * The total size of fadump reserved memory covers for boot memory size
75 * + cpu data size + hpte size and metadata.
76 * Initialize only the area equivalent to boot memory size for CMA use.
77 * The remaining portion of fadump reserved memory will be not given
78 * to CMA and pages for those will stay reserved. boot memory size is
79 * aligned per CMA requirement to satisy cma_init_reserved_mem() call.
80 * But for some reason even if it fails we still have the memory reservation
81 * with us and we can still continue doing fadump.
83 static int __init fadump_cma_init(void)
85 unsigned long long base, size;
88 if (!fw_dump.fadump_enabled)
92 * Do not use CMA if user has provided fadump=nocma kernel parameter.
93 * Return 1 to continue with fadump old behaviour.
98 base = fw_dump.reserve_dump_area_start;
99 size = fw_dump.boot_memory_size;
104 rc = cma_init_reserved_mem(base, size, 0, "fadump_cma", &fadump_cma);
106 pr_err("Failed to init cma area for firmware-assisted dump,%d\n", rc);
108 * Though the CMA init has failed we still have memory
109 * reservation with us. The reserved memory will be
110 * blocked from production system usage. Hence return 1,
111 * so that we can continue with fadump.
117 * If CMA activation fails, keep the pages reserved, instead of
118 * exposing them to buddy allocator. Same as 'fadump=nocma' case.
120 cma_reserve_pages_on_error(fadump_cma);
123 * So we now have successfully initialized cma area for fadump.
125 pr_info("Initialized 0x%lx bytes cma area at %ldMB from 0x%lx "
126 "bytes of memory reserved for firmware-assisted dump\n",
127 cma_get_size(fadump_cma),
128 (unsigned long)cma_get_base(fadump_cma) >> 20,
129 fw_dump.reserve_dump_area_size);
133 static int __init fadump_cma_init(void) { return 1; }
134 #endif /* CONFIG_CMA */
136 /* Scan the Firmware Assisted dump configuration details. */
137 int __init early_init_dt_scan_fw_dump(unsigned long node, const char *uname,
138 int depth, void *data)
141 early_init_dt_scan_reserved_ranges(node);
148 if (strcmp(uname, "rtas") == 0) {
149 rtas_fadump_dt_scan(&fw_dump, node);
153 if (strcmp(uname, "ibm,opal") == 0) {
154 opal_fadump_dt_scan(&fw_dump, node);
162 * If fadump is registered, check if the memory provided
163 * falls within boot memory area and reserved memory area.
165 int is_fadump_memory_area(u64 addr, unsigned long size)
169 if (!fw_dump.dump_registered)
175 d_start = fw_dump.reserve_dump_area_start;
176 d_end = d_start + fw_dump.reserve_dump_area_size;
177 if (((addr + size) > d_start) && (addr <= d_end))
180 return (addr <= fw_dump.boot_mem_top);
183 int should_fadump_crash(void)
185 if (!fw_dump.dump_registered || !fw_dump.fadumphdr_addr)
190 int is_fadump_active(void)
192 return fw_dump.dump_active;
196 * Returns true, if there are no holes in memory area between d_start to d_end,
199 static bool is_fadump_mem_area_contiguous(u64 d_start, u64 d_end)
201 phys_addr_t reg_start, reg_end;
205 for_each_mem_range(i, ®_start, ®_end) {
206 start = max_t(u64, d_start, reg_start);
207 end = min_t(u64, d_end, reg_end);
209 /* Memory hole from d_start to start */
226 * Returns true, if there are no holes in boot memory area,
229 bool is_fadump_boot_mem_contiguous(void)
231 unsigned long d_start, d_end;
235 for (i = 0; i < fw_dump.boot_mem_regs_cnt; i++) {
236 d_start = fw_dump.boot_mem_addr[i];
237 d_end = d_start + fw_dump.boot_mem_sz[i];
239 ret = is_fadump_mem_area_contiguous(d_start, d_end);
248 * Returns true, if there are no holes in reserved memory area,
251 bool is_fadump_reserved_mem_contiguous(void)
255 d_start = fw_dump.reserve_dump_area_start;
256 d_end = d_start + fw_dump.reserve_dump_area_size;
257 return is_fadump_mem_area_contiguous(d_start, d_end);
260 /* Print firmware assisted dump configurations for debugging purpose. */
261 static void __init fadump_show_config(void)
265 pr_debug("Support for firmware-assisted dump (fadump): %s\n",
266 (fw_dump.fadump_supported ? "present" : "no support"));
268 if (!fw_dump.fadump_supported)
271 pr_debug("Fadump enabled : %s\n",
272 (fw_dump.fadump_enabled ? "yes" : "no"));
273 pr_debug("Dump Active : %s\n",
274 (fw_dump.dump_active ? "yes" : "no"));
275 pr_debug("Dump section sizes:\n");
276 pr_debug(" CPU state data size: %lx\n", fw_dump.cpu_state_data_size);
277 pr_debug(" HPTE region size : %lx\n", fw_dump.hpte_region_size);
278 pr_debug(" Boot memory size : %lx\n", fw_dump.boot_memory_size);
279 pr_debug(" Boot memory top : %llx\n", fw_dump.boot_mem_top);
280 pr_debug("Boot memory regions cnt: %llx\n", fw_dump.boot_mem_regs_cnt);
281 for (i = 0; i < fw_dump.boot_mem_regs_cnt; i++) {
282 pr_debug("[%03d] base = %llx, size = %llx\n", i,
283 fw_dump.boot_mem_addr[i], fw_dump.boot_mem_sz[i]);
288 * fadump_calculate_reserve_size(): reserve variable boot area 5% of System RAM
290 * Function to find the largest memory size we need to reserve during early
291 * boot process. This will be the size of the memory that is required for a
292 * kernel to boot successfully.
294 * This function has been taken from phyp-assisted dump feature implementation.
296 * returns larger of 256MB or 5% rounded down to multiples of 256MB.
298 * TODO: Come up with better approach to find out more accurate memory size
299 * that is required for a kernel to boot successfully.
302 static __init u64 fadump_calculate_reserve_size(void)
304 u64 base, size, bootmem_min;
307 if (fw_dump.reserve_bootvar)
308 pr_warn("'fadump_reserve_mem=' parameter is deprecated in favor of 'crashkernel=' parameter.\n");
311 * Check if the size is specified through crashkernel= cmdline
312 * option. If yes, then use that but ignore base as fadump reserves
313 * memory at a predefined offset.
315 ret = parse_crashkernel(boot_command_line, memblock_phys_mem_size(),
317 if (ret == 0 && size > 0) {
318 unsigned long max_size;
320 if (fw_dump.reserve_bootvar)
321 pr_info("Using 'crashkernel=' parameter for memory reservation.\n");
323 fw_dump.reserve_bootvar = (unsigned long)size;
326 * Adjust if the boot memory size specified is above
329 max_size = memblock_phys_mem_size() / MAX_BOOT_MEM_RATIO;
330 if (fw_dump.reserve_bootvar > max_size) {
331 fw_dump.reserve_bootvar = max_size;
332 pr_info("Adjusted boot memory size to %luMB\n",
333 (fw_dump.reserve_bootvar >> 20));
336 return fw_dump.reserve_bootvar;
337 } else if (fw_dump.reserve_bootvar) {
339 * 'fadump_reserve_mem=' is being used to reserve memory
340 * for firmware-assisted dump.
342 return fw_dump.reserve_bootvar;
345 /* divide by 20 to get 5% of value */
346 size = memblock_phys_mem_size() / 20;
348 /* round it down in multiples of 256 */
349 size = size & ~0x0FFFFFFFUL;
351 /* Truncate to memory_limit. We don't want to over reserve the memory.*/
352 if (memory_limit && size > memory_limit)
355 bootmem_min = fw_dump.ops->fadump_get_bootmem_min();
356 return (size > bootmem_min ? size : bootmem_min);
360 * Calculate the total memory size required to be reserved for
361 * firmware-assisted dump registration.
363 static unsigned long __init get_fadump_area_size(void)
365 unsigned long size = 0;
367 size += fw_dump.cpu_state_data_size;
368 size += fw_dump.hpte_region_size;
370 * Account for pagesize alignment of boot memory area destination address.
371 * This faciliates in mmap reading of first kernel's memory.
373 size = PAGE_ALIGN(size);
374 size += fw_dump.boot_memory_size;
375 size += sizeof(struct fadump_crash_info_header);
376 size += sizeof(struct elfhdr); /* ELF core header.*/
377 size += sizeof(struct elf_phdr); /* place holder for cpu notes */
378 /* Program headers for crash memory regions. */
379 size += sizeof(struct elf_phdr) * (memblock_num_regions(memory) + 2);
381 size = PAGE_ALIGN(size);
383 /* This is to hold kernel metadata on platforms that support it */
384 size += (fw_dump.ops->fadump_get_metadata_size ?
385 fw_dump.ops->fadump_get_metadata_size() : 0);
389 static int __init add_boot_mem_region(unsigned long rstart,
392 int i = fw_dump.boot_mem_regs_cnt++;
394 if (fw_dump.boot_mem_regs_cnt > FADUMP_MAX_MEM_REGS) {
395 fw_dump.boot_mem_regs_cnt = FADUMP_MAX_MEM_REGS;
399 pr_debug("Added boot memory range[%d] [%#016lx-%#016lx)\n",
400 i, rstart, (rstart + rsize));
401 fw_dump.boot_mem_addr[i] = rstart;
402 fw_dump.boot_mem_sz[i] = rsize;
407 * Firmware usually has a hard limit on the data it can copy per region.
408 * Honour that by splitting a memory range into multiple regions.
410 static int __init add_boot_mem_regions(unsigned long mstart,
413 unsigned long rstart, rsize, max_size;
417 max_size = fw_dump.max_copy_size ? fw_dump.max_copy_size : msize;
419 if (msize > max_size)
424 ret = add_boot_mem_region(rstart, rsize);
435 static int __init fadump_get_boot_mem_regions(void)
437 unsigned long size, cur_size, hole_size, last_end;
438 unsigned long mem_size = fw_dump.boot_memory_size;
439 phys_addr_t reg_start, reg_end;
443 fw_dump.boot_mem_regs_cnt = 0;
448 for_each_mem_range(i, ®_start, ®_end) {
449 size = reg_end - reg_start;
450 hole_size += (reg_start - last_end);
452 if ((cur_size + size) >= mem_size) {
453 size = (mem_size - cur_size);
454 ret = add_boot_mem_regions(reg_start, size);
460 ret = add_boot_mem_regions(reg_start, size);
466 fw_dump.boot_mem_top = PAGE_ALIGN(fw_dump.boot_memory_size + hole_size);
472 * Returns true, if the given range overlaps with reserved memory ranges
473 * starting at idx. Also, updates idx to index of overlapping memory range
474 * with the given memory range.
477 static bool __init overlaps_reserved_ranges(u64 base, u64 end, int *idx)
482 for (i = *idx; i < reserved_mrange_info.mem_range_cnt; i++) {
483 u64 rbase = reserved_mrange_info.mem_ranges[i].base;
484 u64 rend = rbase + reserved_mrange_info.mem_ranges[i].size;
489 if ((end > rbase) && (base < rend)) {
500 * Locate a suitable memory area to reserve memory for FADump. While at it,
501 * lookup reserved-ranges & avoid overlap with them, as they are used by F/W.
503 static u64 __init fadump_locate_reserve_mem(u64 base, u64 size)
505 struct fadump_memory_range *mrngs;
506 phys_addr_t mstart, mend;
510 mrngs = reserved_mrange_info.mem_ranges;
511 for_each_free_mem_range(i, NUMA_NO_NODE, MEMBLOCK_NONE,
512 &mstart, &mend, NULL) {
513 pr_debug("%llu) mstart: %llx, mend: %llx, base: %llx\n",
514 i, mstart, mend, base);
517 base = PAGE_ALIGN(mstart);
519 while ((mend > base) && ((mend - base) >= size)) {
520 if (!overlaps_reserved_ranges(base, base+size, &idx)) {
525 base = mrngs[idx].base + mrngs[idx].size;
526 base = PAGE_ALIGN(base);
534 int __init fadump_reserve_mem(void)
536 u64 base, size, mem_boundary, bootmem_min;
539 if (!fw_dump.fadump_enabled)
542 if (!fw_dump.fadump_supported) {
543 pr_info("Firmware-Assisted Dump is not supported on this hardware\n");
548 * Initialize boot memory size
549 * If dump is active then we have already calculated the size during
552 if (!fw_dump.dump_active) {
553 fw_dump.boot_memory_size =
554 PAGE_ALIGN(fadump_calculate_reserve_size());
556 if (!fw_dump.nocma) {
557 fw_dump.boot_memory_size =
558 ALIGN(fw_dump.boot_memory_size,
559 CMA_MIN_ALIGNMENT_BYTES);
563 bootmem_min = fw_dump.ops->fadump_get_bootmem_min();
564 if (fw_dump.boot_memory_size < bootmem_min) {
565 pr_err("Can't enable fadump with boot memory size (0x%lx) less than 0x%llx\n",
566 fw_dump.boot_memory_size, bootmem_min);
570 if (!fadump_get_boot_mem_regions()) {
571 pr_err("Too many holes in boot memory area to enable fadump\n");
577 * Calculate the memory boundary.
578 * If memory_limit is less than actual memory boundary then reserve
579 * the memory for fadump beyond the memory_limit and adjust the
580 * memory_limit accordingly, so that the running kernel can run with
581 * specified memory_limit.
583 if (memory_limit && memory_limit < memblock_end_of_DRAM()) {
584 size = get_fadump_area_size();
585 if ((memory_limit + size) < memblock_end_of_DRAM())
586 memory_limit += size;
588 memory_limit = memblock_end_of_DRAM();
589 printk(KERN_INFO "Adjusted memory_limit for firmware-assisted"
590 " dump, now %#016llx\n", memory_limit);
593 mem_boundary = memory_limit;
595 mem_boundary = memblock_end_of_DRAM();
597 base = fw_dump.boot_mem_top;
598 size = get_fadump_area_size();
599 fw_dump.reserve_dump_area_size = size;
600 if (fw_dump.dump_active) {
601 pr_info("Firmware-assisted dump is active.\n");
603 #ifdef CONFIG_HUGETLB_PAGE
605 * FADump capture kernel doesn't care much about hugepages.
606 * In fact, handling hugepages in capture kernel is asking for
607 * trouble. So, disable HugeTLB support when fadump is active.
609 hugetlb_disabled = true;
612 * If last boot has crashed then reserve all the memory
613 * above boot memory size so that we don't touch it until
614 * dump is written to disk by userspace tool. This memory
615 * can be released for general use by invalidating fadump.
617 fadump_reserve_crash_area(base);
619 pr_debug("fadumphdr_addr = %#016lx\n", fw_dump.fadumphdr_addr);
620 pr_debug("Reserve dump area start address: 0x%lx\n",
621 fw_dump.reserve_dump_area_start);
624 * Reserve memory at an offset closer to bottom of the RAM to
625 * minimize the impact of memory hot-remove operation.
627 base = fadump_locate_reserve_mem(base, size);
629 if (!base || (base + size > mem_boundary)) {
630 pr_err("Failed to find memory chunk for reservation!\n");
633 fw_dump.reserve_dump_area_start = base;
636 * Calculate the kernel metadata address and register it with
637 * f/w if the platform supports.
639 if (fw_dump.ops->fadump_setup_metadata &&
640 (fw_dump.ops->fadump_setup_metadata(&fw_dump) < 0))
643 if (memblock_reserve(base, size)) {
644 pr_err("Failed to reserve memory!\n");
648 pr_info("Reserved %lldMB of memory at %#016llx (System RAM: %lldMB)\n",
649 (size >> 20), base, (memblock_phys_mem_size() >> 20));
651 ret = fadump_cma_init();
656 fw_dump.fadump_enabled = 0;
657 fw_dump.reserve_dump_area_size = 0;
661 /* Look for fadump= cmdline option. */
662 static int __init early_fadump_param(char *p)
667 if (strncmp(p, "on", 2) == 0)
668 fw_dump.fadump_enabled = 1;
669 else if (strncmp(p, "off", 3) == 0)
670 fw_dump.fadump_enabled = 0;
671 else if (strncmp(p, "nocma", 5) == 0) {
672 fw_dump.fadump_enabled = 1;
678 early_param("fadump", early_fadump_param);
681 * Look for fadump_reserve_mem= cmdline option
682 * TODO: Remove references to 'fadump_reserve_mem=' parameter,
683 * the sooner 'crashkernel=' parameter is accustomed to.
685 static int __init early_fadump_reserve_mem(char *p)
688 fw_dump.reserve_bootvar = memparse(p, &p);
691 early_param("fadump_reserve_mem", early_fadump_reserve_mem);
693 void crash_fadump(struct pt_regs *regs, const char *str)
696 struct fadump_crash_info_header *fdh = NULL;
697 int old_cpu, this_cpu;
698 /* Do not include first CPU */
699 unsigned int ncpus = num_online_cpus() - 1;
701 if (!should_fadump_crash())
705 * old_cpu == -1 means this is the first CPU which has come here,
706 * go ahead and trigger fadump.
708 * old_cpu != -1 means some other CPU has already on it's way
709 * to trigger fadump, just keep looping here.
711 this_cpu = smp_processor_id();
712 old_cpu = cmpxchg(&crashing_cpu, -1, this_cpu);
715 atomic_inc(&cpus_in_fadump);
718 * We can't loop here indefinitely. Wait as long as fadump
719 * is in force. If we race with fadump un-registration this
720 * loop will break and then we go down to normal panic path
721 * and reboot. If fadump is in force the first crashing
722 * cpu will definitely trigger fadump.
724 while (fw_dump.dump_registered)
729 fdh = __va(fw_dump.fadumphdr_addr);
730 fdh->crashing_cpu = crashing_cpu;
731 crash_save_vmcoreinfo();
736 ppc_save_regs(&fdh->regs);
738 fdh->cpu_mask = *cpu_online_mask;
741 * If we came in via system reset, wait a while for the secondary
744 if (TRAP(&(fdh->regs)) == INTERRUPT_SYSTEM_RESET) {
745 msecs = CRASH_TIMEOUT;
746 while ((atomic_read(&cpus_in_fadump) < ncpus) && (--msecs > 0))
750 fw_dump.ops->fadump_trigger(fdh, str);
753 u32 *__init fadump_regs_to_elf_notes(u32 *buf, struct pt_regs *regs)
755 struct elf_prstatus prstatus;
757 memset(&prstatus, 0, sizeof(prstatus));
759 * FIXME: How do i get PID? Do I really need it?
760 * prstatus.pr_pid = ????
762 elf_core_copy_regs(&prstatus.pr_reg, regs);
763 buf = append_elf_note(buf, CRASH_CORE_NOTE_NAME, NT_PRSTATUS,
764 &prstatus, sizeof(prstatus));
768 void __init fadump_update_elfcore_header(char *bufp)
770 struct elf_phdr *phdr;
772 bufp += sizeof(struct elfhdr);
774 /* First note is a place holder for cpu notes info. */
775 phdr = (struct elf_phdr *)bufp;
777 if (phdr->p_type == PT_NOTE) {
778 phdr->p_paddr = __pa(fw_dump.cpu_notes_buf_vaddr);
779 phdr->p_offset = phdr->p_paddr;
780 phdr->p_filesz = fw_dump.cpu_notes_buf_size;
781 phdr->p_memsz = fw_dump.cpu_notes_buf_size;
786 static void *__init fadump_alloc_buffer(unsigned long size)
788 unsigned long count, i;
792 vaddr = alloc_pages_exact(size, GFP_KERNEL | __GFP_ZERO);
796 count = PAGE_ALIGN(size) / PAGE_SIZE;
797 page = virt_to_page(vaddr);
798 for (i = 0; i < count; i++)
799 mark_page_reserved(page + i);
803 static void fadump_free_buffer(unsigned long vaddr, unsigned long size)
805 free_reserved_area((void *)vaddr, (void *)(vaddr + size), -1, NULL);
808 s32 __init fadump_setup_cpu_notes_buf(u32 num_cpus)
810 /* Allocate buffer to hold cpu crash notes. */
811 fw_dump.cpu_notes_buf_size = num_cpus * sizeof(note_buf_t);
812 fw_dump.cpu_notes_buf_size = PAGE_ALIGN(fw_dump.cpu_notes_buf_size);
813 fw_dump.cpu_notes_buf_vaddr =
814 (unsigned long)fadump_alloc_buffer(fw_dump.cpu_notes_buf_size);
815 if (!fw_dump.cpu_notes_buf_vaddr) {
816 pr_err("Failed to allocate %ld bytes for CPU notes buffer\n",
817 fw_dump.cpu_notes_buf_size);
821 pr_debug("Allocated buffer for cpu notes of size %ld at 0x%lx\n",
822 fw_dump.cpu_notes_buf_size,
823 fw_dump.cpu_notes_buf_vaddr);
827 void fadump_free_cpu_notes_buf(void)
829 if (!fw_dump.cpu_notes_buf_vaddr)
832 fadump_free_buffer(fw_dump.cpu_notes_buf_vaddr,
833 fw_dump.cpu_notes_buf_size);
834 fw_dump.cpu_notes_buf_vaddr = 0;
835 fw_dump.cpu_notes_buf_size = 0;
838 static void fadump_free_mem_ranges(struct fadump_mrange_info *mrange_info)
840 if (mrange_info->is_static) {
841 mrange_info->mem_range_cnt = 0;
845 kfree(mrange_info->mem_ranges);
846 memset((void *)((u64)mrange_info + RNG_NAME_SZ), 0,
847 (sizeof(struct fadump_mrange_info) - RNG_NAME_SZ));
851 * Allocate or reallocate mem_ranges array in incremental units
854 static int fadump_alloc_mem_ranges(struct fadump_mrange_info *mrange_info)
856 struct fadump_memory_range *new_array;
859 new_size = mrange_info->mem_ranges_sz + PAGE_SIZE;
860 pr_debug("Allocating %llu bytes of memory for %s memory ranges\n",
861 new_size, mrange_info->name);
863 new_array = krealloc(mrange_info->mem_ranges, new_size, GFP_KERNEL);
864 if (new_array == NULL) {
865 pr_err("Insufficient memory for setting up %s memory ranges\n",
867 fadump_free_mem_ranges(mrange_info);
871 mrange_info->mem_ranges = new_array;
872 mrange_info->mem_ranges_sz = new_size;
873 mrange_info->max_mem_ranges = (new_size /
874 sizeof(struct fadump_memory_range));
877 static inline int fadump_add_mem_range(struct fadump_mrange_info *mrange_info,
880 struct fadump_memory_range *mem_ranges = mrange_info->mem_ranges;
881 bool is_adjacent = false;
888 * Fold adjacent memory ranges to bring down the memory ranges/
889 * PT_LOAD segments count.
891 if (mrange_info->mem_range_cnt) {
892 start = mem_ranges[mrange_info->mem_range_cnt - 1].base;
893 size = mem_ranges[mrange_info->mem_range_cnt - 1].size;
896 * Boot memory area needs separate PT_LOAD segment(s) as it
897 * is moved to a different location at the time of crash.
898 * So, fold only if the region is not boot memory area.
900 if ((start + size) == base && start >= fw_dump.boot_mem_top)
904 /* resize the array on reaching the limit */
905 if (mrange_info->mem_range_cnt == mrange_info->max_mem_ranges) {
908 if (mrange_info->is_static) {
909 pr_err("Reached array size limit for %s memory ranges\n",
914 ret = fadump_alloc_mem_ranges(mrange_info);
918 /* Update to the new resized array */
919 mem_ranges = mrange_info->mem_ranges;
923 mem_ranges[mrange_info->mem_range_cnt].base = start;
924 mrange_info->mem_range_cnt++;
927 mem_ranges[mrange_info->mem_range_cnt - 1].size = (end - start);
928 pr_debug("%s_memory_range[%d] [%#016llx-%#016llx], %#llx bytes\n",
929 mrange_info->name, (mrange_info->mem_range_cnt - 1),
930 start, end - 1, (end - start));
934 static int fadump_exclude_reserved_area(u64 start, u64 end)
936 u64 ra_start, ra_end;
939 ra_start = fw_dump.reserve_dump_area_start;
940 ra_end = ra_start + fw_dump.reserve_dump_area_size;
942 if ((ra_start < end) && (ra_end > start)) {
943 if ((start < ra_start) && (end > ra_end)) {
944 ret = fadump_add_mem_range(&crash_mrange_info,
949 ret = fadump_add_mem_range(&crash_mrange_info,
951 } else if (start < ra_start) {
952 ret = fadump_add_mem_range(&crash_mrange_info,
954 } else if (ra_end < end) {
955 ret = fadump_add_mem_range(&crash_mrange_info,
959 ret = fadump_add_mem_range(&crash_mrange_info, start, end);
964 static int fadump_init_elfcore_header(char *bufp)
968 elf = (struct elfhdr *) bufp;
969 bufp += sizeof(struct elfhdr);
970 memcpy(elf->e_ident, ELFMAG, SELFMAG);
971 elf->e_ident[EI_CLASS] = ELF_CLASS;
972 elf->e_ident[EI_DATA] = ELF_DATA;
973 elf->e_ident[EI_VERSION] = EV_CURRENT;
974 elf->e_ident[EI_OSABI] = ELF_OSABI;
975 memset(elf->e_ident+EI_PAD, 0, EI_NIDENT-EI_PAD);
976 elf->e_type = ET_CORE;
977 elf->e_machine = ELF_ARCH;
978 elf->e_version = EV_CURRENT;
980 elf->e_phoff = sizeof(struct elfhdr);
983 if (IS_ENABLED(CONFIG_PPC64_ELF_ABI_V2))
985 else if (IS_ENABLED(CONFIG_PPC64_ELF_ABI_V1))
990 elf->e_ehsize = sizeof(struct elfhdr);
991 elf->e_phentsize = sizeof(struct elf_phdr);
993 elf->e_shentsize = 0;
1001 * Traverse through memblock structure and setup crash memory ranges. These
1002 * ranges will be used create PT_LOAD program headers in elfcore header.
1004 static int fadump_setup_crash_memory_ranges(void)
1009 pr_debug("Setup crash memory ranges.\n");
1010 crash_mrange_info.mem_range_cnt = 0;
1013 * Boot memory region(s) registered with firmware are moved to
1014 * different location at the time of crash. Create separate program
1015 * header(s) for this memory chunk(s) with the correct offset.
1017 for (i = 0; i < fw_dump.boot_mem_regs_cnt; i++) {
1018 start = fw_dump.boot_mem_addr[i];
1019 end = start + fw_dump.boot_mem_sz[i];
1020 ret = fadump_add_mem_range(&crash_mrange_info, start, end);
1025 for_each_mem_range(i, &start, &end) {
1027 * skip the memory chunk that is already added
1028 * (0 through boot_memory_top).
1030 if (start < fw_dump.boot_mem_top) {
1031 if (end > fw_dump.boot_mem_top)
1032 start = fw_dump.boot_mem_top;
1037 /* add this range excluding the reserved dump area. */
1038 ret = fadump_exclude_reserved_area(start, end);
1047 * If the given physical address falls within the boot memory region then
1048 * return the relocated address that points to the dump region reserved
1049 * for saving initial boot memory contents.
1051 static inline unsigned long fadump_relocate(unsigned long paddr)
1053 unsigned long raddr, rstart, rend, rlast, hole_size;
1059 for (i = 0; i < fw_dump.boot_mem_regs_cnt; i++) {
1060 rstart = fw_dump.boot_mem_addr[i];
1061 rend = rstart + fw_dump.boot_mem_sz[i];
1062 hole_size += (rstart - rlast);
1064 if (paddr >= rstart && paddr < rend) {
1065 raddr += fw_dump.boot_mem_dest_addr - hole_size;
1072 pr_debug("vmcoreinfo: paddr = 0x%lx, raddr = 0x%lx\n", paddr, raddr);
1076 static int fadump_create_elfcore_headers(char *bufp)
1078 unsigned long long raddr, offset;
1079 struct elf_phdr *phdr;
1083 fadump_init_elfcore_header(bufp);
1084 elf = (struct elfhdr *)bufp;
1085 bufp += sizeof(struct elfhdr);
1088 * setup ELF PT_NOTE, place holder for cpu notes info. The notes info
1089 * will be populated during second kernel boot after crash. Hence
1090 * this PT_NOTE will always be the first elf note.
1092 * NOTE: Any new ELF note addition should be placed after this note.
1094 phdr = (struct elf_phdr *)bufp;
1095 bufp += sizeof(struct elf_phdr);
1096 phdr->p_type = PT_NOTE;
1108 /* setup ELF PT_NOTE for vmcoreinfo */
1109 phdr = (struct elf_phdr *)bufp;
1110 bufp += sizeof(struct elf_phdr);
1111 phdr->p_type = PT_NOTE;
1116 phdr->p_paddr = fadump_relocate(paddr_vmcoreinfo_note());
1117 phdr->p_offset = phdr->p_paddr;
1118 phdr->p_memsz = phdr->p_filesz = VMCOREINFO_NOTE_SIZE;
1120 /* Increment number of program headers. */
1123 /* setup PT_LOAD sections. */
1126 raddr = fw_dump.boot_mem_addr[0];
1127 for (i = 0; i < crash_mrange_info.mem_range_cnt; i++) {
1130 mbase = crash_mrange_info.mem_ranges[i].base;
1131 msize = crash_mrange_info.mem_ranges[i].size;
1135 phdr = (struct elf_phdr *)bufp;
1136 bufp += sizeof(struct elf_phdr);
1137 phdr->p_type = PT_LOAD;
1138 phdr->p_flags = PF_R|PF_W|PF_X;
1139 phdr->p_offset = mbase;
1141 if (mbase == raddr) {
1143 * The entire real memory region will be moved by
1144 * firmware to the specified destination_address.
1145 * Hence set the correct offset.
1147 phdr->p_offset = fw_dump.boot_mem_dest_addr + offset;
1148 if (j < (fw_dump.boot_mem_regs_cnt - 1)) {
1149 offset += fw_dump.boot_mem_sz[j];
1150 raddr = fw_dump.boot_mem_addr[++j];
1154 phdr->p_paddr = mbase;
1155 phdr->p_vaddr = (unsigned long)__va(mbase);
1156 phdr->p_filesz = msize;
1157 phdr->p_memsz = msize;
1160 /* Increment number of program headers. */
1166 static unsigned long init_fadump_header(unsigned long addr)
1168 struct fadump_crash_info_header *fdh;
1174 addr += sizeof(struct fadump_crash_info_header);
1176 memset(fdh, 0, sizeof(struct fadump_crash_info_header));
1177 fdh->magic_number = FADUMP_CRASH_INFO_MAGIC;
1178 fdh->elfcorehdr_addr = addr;
1179 /* We will set the crashing cpu id in crash_fadump() during crash. */
1180 fdh->crashing_cpu = FADUMP_CPU_UNKNOWN;
1182 * When LPAR is terminated by PYHP, ensure all possible CPUs'
1183 * register data is processed while exporting the vmcore.
1185 fdh->cpu_mask = *cpu_possible_mask;
1190 static int register_fadump(void)
1197 * If no memory is reserved then we can not register for firmware-
1200 if (!fw_dump.reserve_dump_area_size)
1203 ret = fadump_setup_crash_memory_ranges();
1207 addr = fw_dump.fadumphdr_addr;
1209 /* Initialize fadump crash info header. */
1210 addr = init_fadump_header(addr);
1213 pr_debug("Creating ELF core headers at %#016lx\n", addr);
1214 fadump_create_elfcore_headers(vaddr);
1216 /* register the future kernel dump with firmware. */
1217 pr_debug("Registering for firmware-assisted kernel dump...\n");
1218 return fw_dump.ops->fadump_register(&fw_dump);
1221 void fadump_cleanup(void)
1223 if (!fw_dump.fadump_supported)
1226 /* Invalidate the registration only if dump is active. */
1227 if (fw_dump.dump_active) {
1228 pr_debug("Invalidating firmware-assisted dump registration\n");
1229 fw_dump.ops->fadump_invalidate(&fw_dump);
1230 } else if (fw_dump.dump_registered) {
1231 /* Un-register Firmware-assisted dump if it was registered. */
1232 fw_dump.ops->fadump_unregister(&fw_dump);
1233 fadump_free_mem_ranges(&crash_mrange_info);
1236 if (fw_dump.ops->fadump_cleanup)
1237 fw_dump.ops->fadump_cleanup(&fw_dump);
1240 static void fadump_free_reserved_memory(unsigned long start_pfn,
1241 unsigned long end_pfn)
1244 unsigned long time_limit = jiffies + HZ;
1246 pr_info("freeing reserved memory (0x%llx - 0x%llx)\n",
1247 PFN_PHYS(start_pfn), PFN_PHYS(end_pfn));
1249 for (pfn = start_pfn; pfn < end_pfn; pfn++) {
1250 free_reserved_page(pfn_to_page(pfn));
1252 if (time_after(jiffies, time_limit)) {
1254 time_limit = jiffies + HZ;
1260 * Skip memory holes and free memory that was actually reserved.
1262 static void fadump_release_reserved_area(u64 start, u64 end)
1264 unsigned long reg_spfn, reg_epfn;
1265 u64 tstart, tend, spfn, epfn;
1268 spfn = PHYS_PFN(start);
1269 epfn = PHYS_PFN(end);
1271 for_each_mem_pfn_range(i, MAX_NUMNODES, ®_spfn, ®_epfn, NULL) {
1272 tstart = max_t(u64, spfn, reg_spfn);
1273 tend = min_t(u64, epfn, reg_epfn);
1275 if (tstart < tend) {
1276 fadump_free_reserved_memory(tstart, tend);
1287 * Sort the mem ranges in-place and merge adjacent ranges
1288 * to minimize the memory ranges count.
1290 static void sort_and_merge_mem_ranges(struct fadump_mrange_info *mrange_info)
1292 struct fadump_memory_range *mem_ranges;
1296 if (!reserved_mrange_info.mem_range_cnt)
1299 /* Sort the memory ranges */
1300 mem_ranges = mrange_info->mem_ranges;
1301 for (i = 0; i < mrange_info->mem_range_cnt; i++) {
1303 for (j = (i + 1); j < mrange_info->mem_range_cnt; j++) {
1304 if (mem_ranges[idx].base > mem_ranges[j].base)
1308 swap(mem_ranges[idx], mem_ranges[i]);
1311 /* Merge adjacent reserved ranges */
1313 for (i = 1; i < mrange_info->mem_range_cnt; i++) {
1314 base = mem_ranges[i-1].base;
1315 size = mem_ranges[i-1].size;
1316 if (mem_ranges[i].base == (base + size))
1317 mem_ranges[idx].size += mem_ranges[i].size;
1323 mem_ranges[idx] = mem_ranges[i];
1326 mrange_info->mem_range_cnt = idx + 1;
1330 * Scan reserved-ranges to consider them while reserving/releasing
1331 * memory for FADump.
1333 static void __init early_init_dt_scan_reserved_ranges(unsigned long node)
1339 /* reserved-ranges already scanned */
1340 if (reserved_mrange_info.mem_range_cnt != 0)
1343 prop = of_get_flat_dt_prop(node, "reserved-ranges", &len);
1348 * Each reserved range is an (address,size) pair, 2 cells each,
1349 * totalling 4 cells per range.
1351 for (i = 0; i < len / (sizeof(*prop) * 4); i++) {
1354 base = of_read_number(prop + (i * 4) + 0, 2);
1355 size = of_read_number(prop + (i * 4) + 2, 2);
1358 ret = fadump_add_mem_range(&reserved_mrange_info,
1361 pr_warn("some reserved ranges are ignored!\n");
1367 /* Compact reserved ranges */
1368 sort_and_merge_mem_ranges(&reserved_mrange_info);
1372 * Release the memory that was reserved during early boot to preserve the
1373 * crash'ed kernel's memory contents except reserved dump area (permanent
1374 * reservation) and reserved ranges used by F/W. The released memory will
1375 * be available for general use.
1377 static void fadump_release_memory(u64 begin, u64 end)
1379 u64 ra_start, ra_end, tstart;
1382 ra_start = fw_dump.reserve_dump_area_start;
1383 ra_end = ra_start + fw_dump.reserve_dump_area_size;
1386 * If reserved ranges array limit is hit, overwrite the last reserved
1387 * memory range with reserved dump area to ensure it is excluded from
1388 * the memory being released (reused for next FADump registration).
1390 if (reserved_mrange_info.mem_range_cnt ==
1391 reserved_mrange_info.max_mem_ranges)
1392 reserved_mrange_info.mem_range_cnt--;
1394 ret = fadump_add_mem_range(&reserved_mrange_info, ra_start, ra_end);
1398 /* Get the reserved ranges list in order first. */
1399 sort_and_merge_mem_ranges(&reserved_mrange_info);
1401 /* Exclude reserved ranges and release remaining memory */
1403 for (i = 0; i < reserved_mrange_info.mem_range_cnt; i++) {
1404 ra_start = reserved_mrange_info.mem_ranges[i].base;
1405 ra_end = ra_start + reserved_mrange_info.mem_ranges[i].size;
1407 if (tstart >= ra_end)
1410 if (tstart < ra_start)
1411 fadump_release_reserved_area(tstart, ra_start);
1416 fadump_release_reserved_area(tstart, end);
1419 static void fadump_invalidate_release_mem(void)
1421 mutex_lock(&fadump_mutex);
1422 if (!fw_dump.dump_active) {
1423 mutex_unlock(&fadump_mutex);
1428 mutex_unlock(&fadump_mutex);
1430 fadump_release_memory(fw_dump.boot_mem_top, memblock_end_of_DRAM());
1431 fadump_free_cpu_notes_buf();
1434 * Setup kernel metadata and initialize the kernel dump
1435 * memory structure for FADump re-registration.
1437 if (fw_dump.ops->fadump_setup_metadata &&
1438 (fw_dump.ops->fadump_setup_metadata(&fw_dump) < 0))
1439 pr_warn("Failed to setup kernel metadata!\n");
1440 fw_dump.ops->fadump_init_mem_struct(&fw_dump);
1443 static ssize_t release_mem_store(struct kobject *kobj,
1444 struct kobj_attribute *attr,
1445 const char *buf, size_t count)
1449 if (!fw_dump.dump_active)
1452 if (kstrtoint(buf, 0, &input))
1457 * Take away the '/proc/vmcore'. We are releasing the dump
1458 * memory, hence it will not be valid anymore.
1460 #ifdef CONFIG_PROC_VMCORE
1463 fadump_invalidate_release_mem();
1470 /* Release the reserved memory and disable the FADump */
1471 static void __init unregister_fadump(void)
1474 fadump_release_memory(fw_dump.reserve_dump_area_start,
1475 fw_dump.reserve_dump_area_size);
1476 fw_dump.fadump_enabled = 0;
1477 kobject_put(fadump_kobj);
1480 static ssize_t enabled_show(struct kobject *kobj,
1481 struct kobj_attribute *attr,
1484 return sprintf(buf, "%d\n", fw_dump.fadump_enabled);
1487 static ssize_t mem_reserved_show(struct kobject *kobj,
1488 struct kobj_attribute *attr,
1491 return sprintf(buf, "%ld\n", fw_dump.reserve_dump_area_size);
1494 static ssize_t registered_show(struct kobject *kobj,
1495 struct kobj_attribute *attr,
1498 return sprintf(buf, "%d\n", fw_dump.dump_registered);
1501 static ssize_t registered_store(struct kobject *kobj,
1502 struct kobj_attribute *attr,
1503 const char *buf, size_t count)
1508 if (!fw_dump.fadump_enabled || fw_dump.dump_active)
1511 if (kstrtoint(buf, 0, &input))
1514 mutex_lock(&fadump_mutex);
1518 if (fw_dump.dump_registered == 0) {
1522 /* Un-register Firmware-assisted dump */
1523 pr_debug("Un-register firmware-assisted dump\n");
1524 fw_dump.ops->fadump_unregister(&fw_dump);
1527 if (fw_dump.dump_registered == 1) {
1528 /* Un-register Firmware-assisted dump */
1529 fw_dump.ops->fadump_unregister(&fw_dump);
1531 /* Register Firmware-assisted dump */
1532 ret = register_fadump();
1540 mutex_unlock(&fadump_mutex);
1541 return ret < 0 ? ret : count;
1544 static int fadump_region_show(struct seq_file *m, void *private)
1546 if (!fw_dump.fadump_enabled)
1549 mutex_lock(&fadump_mutex);
1550 fw_dump.ops->fadump_region_show(&fw_dump, m);
1551 mutex_unlock(&fadump_mutex);
1555 static struct kobj_attribute release_attr = __ATTR_WO(release_mem);
1556 static struct kobj_attribute enable_attr = __ATTR_RO(enabled);
1557 static struct kobj_attribute register_attr = __ATTR_RW(registered);
1558 static struct kobj_attribute mem_reserved_attr = __ATTR_RO(mem_reserved);
1560 static struct attribute *fadump_attrs[] = {
1562 ®ister_attr.attr,
1563 &mem_reserved_attr.attr,
1567 ATTRIBUTE_GROUPS(fadump);
1569 DEFINE_SHOW_ATTRIBUTE(fadump_region);
1571 static void __init fadump_init_files(void)
1575 fadump_kobj = kobject_create_and_add("fadump", kernel_kobj);
1577 pr_err("failed to create fadump kobject\n");
1581 debugfs_create_file("fadump_region", 0444, arch_debugfs_dir, NULL,
1582 &fadump_region_fops);
1584 if (fw_dump.dump_active) {
1585 rc = sysfs_create_file(fadump_kobj, &release_attr.attr);
1587 pr_err("unable to create release_mem sysfs file (%d)\n",
1591 rc = sysfs_create_groups(fadump_kobj, fadump_groups);
1593 pr_err("sysfs group creation failed (%d), unregistering FADump",
1595 unregister_fadump();
1600 * The FADump sysfs are moved from kernel_kobj to fadump_kobj need to
1601 * create symlink at old location to maintain backward compatibility.
1603 * - fadump_enabled -> fadump/enabled
1604 * - fadump_registered -> fadump/registered
1605 * - fadump_release_mem -> fadump/release_mem
1607 rc = compat_only_sysfs_link_entry_to_kobj(kernel_kobj, fadump_kobj,
1608 "enabled", "fadump_enabled");
1610 pr_err("unable to create fadump_enabled symlink (%d)", rc);
1614 rc = compat_only_sysfs_link_entry_to_kobj(kernel_kobj, fadump_kobj,
1616 "fadump_registered");
1618 pr_err("unable to create fadump_registered symlink (%d)", rc);
1619 sysfs_remove_link(kernel_kobj, "fadump_enabled");
1623 if (fw_dump.dump_active) {
1624 rc = compat_only_sysfs_link_entry_to_kobj(kernel_kobj,
1627 "fadump_release_mem");
1629 pr_err("unable to create fadump_release_mem symlink (%d)",
1636 * Prepare for firmware-assisted dump.
1638 int __init setup_fadump(void)
1640 if (!fw_dump.fadump_supported)
1643 fadump_init_files();
1644 fadump_show_config();
1646 if (!fw_dump.fadump_enabled)
1650 * If dump data is available then see if it is valid and prepare for
1651 * saving it to the disk.
1653 if (fw_dump.dump_active) {
1655 * if dump process fails then invalidate the registration
1656 * and release memory before proceeding for re-registration.
1658 if (fw_dump.ops->fadump_process(&fw_dump) < 0)
1659 fadump_invalidate_release_mem();
1661 /* Initialize the kernel dump memory structure and register with f/w */
1662 else if (fw_dump.reserve_dump_area_size) {
1663 fw_dump.ops->fadump_init_mem_struct(&fw_dump);
1668 * In case of panic, fadump is triggered via ppc_panic_event()
1669 * panic notifier. Setting crash_kexec_post_notifiers to 'true'
1670 * lets panic() function take crash friendly path before panic
1671 * notifiers are invoked.
1673 crash_kexec_post_notifiers = true;
1678 * Use subsys_initcall_sync() here because there is dependency with
1679 * crash_save_vmcoreinfo_init(), which must run first to ensure vmcoreinfo initialization
1680 * is done before registering with f/w.
1682 subsys_initcall_sync(setup_fadump);
1683 #else /* !CONFIG_PRESERVE_FA_DUMP */
1685 /* Scan the Firmware Assisted dump configuration details. */
1686 int __init early_init_dt_scan_fw_dump(unsigned long node, const char *uname,
1687 int depth, void *data)
1689 if ((depth != 1) || (strcmp(uname, "ibm,opal") != 0))
1692 opal_fadump_dt_scan(&fw_dump, node);
1697 * When dump is active but PRESERVE_FA_DUMP is enabled on the kernel,
1698 * preserve crash data. The subsequent memory preserving kernel boot
1699 * is likely to process this crash data.
1701 int __init fadump_reserve_mem(void)
1703 if (fw_dump.dump_active) {
1705 * If last boot has crashed then reserve all the memory
1706 * above boot memory to preserve crash data.
1708 pr_info("Preserving crash data for processing in next boot.\n");
1709 fadump_reserve_crash_area(fw_dump.boot_mem_top);
1711 pr_debug("FADump-aware kernel..\n");
1715 #endif /* CONFIG_PRESERVE_FA_DUMP */
1717 /* Preserve everything above the base address */
1718 static void __init fadump_reserve_crash_area(u64 base)
1720 u64 i, mstart, mend, msize;
1722 for_each_mem_range(i, &mstart, &mend) {
1723 msize = mend - mstart;
1725 if ((mstart + msize) < base)
1728 if (mstart < base) {
1729 msize -= (base - mstart);
1733 pr_info("Reserving %lluMB of memory at %#016llx for preserving crash data",
1734 (msize >> 20), mstart);
1735 memblock_reserve(mstart, msize);
1739 unsigned long __init arch_reserved_kernel_pages(void)
1741 return memblock_reserved_size() / PAGE_SIZE;