]> asedeno.scripts.mit.edu Git - linux.git/blob - arch/x86/kernel/kexec-bzimage64.c
kexec: support for kexec on panic using new system call
[linux.git] / arch / x86 / kernel / kexec-bzimage64.c
1 /*
2  * Kexec bzImage loader
3  *
4  * Copyright (C) 2014 Red Hat Inc.
5  * Authors:
6  *      Vivek Goyal <vgoyal@redhat.com>
7  *
8  * This source code is licensed under the GNU General Public License,
9  * Version 2.  See the file COPYING for more details.
10  */
11
12 #define pr_fmt(fmt)     "kexec-bzImage64: " fmt
13
14 #include <linux/string.h>
15 #include <linux/printk.h>
16 #include <linux/errno.h>
17 #include <linux/slab.h>
18 #include <linux/kexec.h>
19 #include <linux/kernel.h>
20 #include <linux/mm.h>
21
22 #include <asm/bootparam.h>
23 #include <asm/setup.h>
24 #include <asm/crash.h>
25
26 #define MAX_ELFCOREHDR_STR_LEN  30      /* elfcorehdr=0x<64bit-value> */
27
28 /*
29  * Defines lowest physical address for various segments. Not sure where
30  * exactly these limits came from. Current bzimage64 loader in kexec-tools
31  * uses these so I am retaining it. It can be changed over time as we gain
32  * more insight.
33  */
34 #define MIN_PURGATORY_ADDR      0x3000
35 #define MIN_BOOTPARAM_ADDR      0x3000
36 #define MIN_KERNEL_LOAD_ADDR    0x100000
37 #define MIN_INITRD_LOAD_ADDR    0x1000000
38
39 /*
40  * This is a place holder for all boot loader specific data structure which
41  * gets allocated in one call but gets freed much later during cleanup
42  * time. Right now there is only one field but it can grow as need be.
43  */
44 struct bzimage64_data {
45         /*
46          * Temporary buffer to hold bootparams buffer. This should be
47          * freed once the bootparam segment has been loaded.
48          */
49         void *bootparams_buf;
50 };
51
52 static int setup_initrd(struct boot_params *params,
53                 unsigned long initrd_load_addr, unsigned long initrd_len)
54 {
55         params->hdr.ramdisk_image = initrd_load_addr & 0xffffffffUL;
56         params->hdr.ramdisk_size = initrd_len & 0xffffffffUL;
57
58         params->ext_ramdisk_image = initrd_load_addr >> 32;
59         params->ext_ramdisk_size = initrd_len >> 32;
60
61         return 0;
62 }
63
64 static int setup_cmdline(struct kimage *image, struct boot_params *params,
65                          unsigned long bootparams_load_addr,
66                          unsigned long cmdline_offset, char *cmdline,
67                          unsigned long cmdline_len)
68 {
69         char *cmdline_ptr = ((char *)params) + cmdline_offset;
70         unsigned long cmdline_ptr_phys, len;
71         uint32_t cmdline_low_32, cmdline_ext_32;
72
73         memcpy(cmdline_ptr, cmdline, cmdline_len);
74         if (image->type == KEXEC_TYPE_CRASH) {
75                 len = sprintf(cmdline_ptr + cmdline_len - 1,
76                         " elfcorehdr=0x%lx", image->arch.elf_load_addr);
77                 cmdline_len += len;
78         }
79         cmdline_ptr[cmdline_len - 1] = '\0';
80
81         pr_debug("Final command line is: %s\n", cmdline_ptr);
82         cmdline_ptr_phys = bootparams_load_addr + cmdline_offset;
83         cmdline_low_32 = cmdline_ptr_phys & 0xffffffffUL;
84         cmdline_ext_32 = cmdline_ptr_phys >> 32;
85
86         params->hdr.cmd_line_ptr = cmdline_low_32;
87         if (cmdline_ext_32)
88                 params->ext_cmd_line_ptr = cmdline_ext_32;
89
90         return 0;
91 }
92
93 static int setup_memory_map_entries(struct boot_params *params)
94 {
95         unsigned int nr_e820_entries;
96
97         nr_e820_entries = e820_saved.nr_map;
98
99         /* TODO: Pass entries more than E820MAX in bootparams setup data */
100         if (nr_e820_entries > E820MAX)
101                 nr_e820_entries = E820MAX;
102
103         params->e820_entries = nr_e820_entries;
104         memcpy(&params->e820_map, &e820_saved.map,
105                nr_e820_entries * sizeof(struct e820entry));
106
107         return 0;
108 }
109
110 static int setup_boot_parameters(struct kimage *image,
111                                  struct boot_params *params)
112 {
113         unsigned int nr_e820_entries;
114         unsigned long long mem_k, start, end;
115         int i, ret = 0;
116
117         /* Get subarch from existing bootparams */
118         params->hdr.hardware_subarch = boot_params.hdr.hardware_subarch;
119
120         /* Copying screen_info will do? */
121         memcpy(&params->screen_info, &boot_params.screen_info,
122                                 sizeof(struct screen_info));
123
124         /* Fill in memsize later */
125         params->screen_info.ext_mem_k = 0;
126         params->alt_mem_k = 0;
127
128         /* Default APM info */
129         memset(&params->apm_bios_info, 0, sizeof(params->apm_bios_info));
130
131         /* Default drive info */
132         memset(&params->hd0_info, 0, sizeof(params->hd0_info));
133         memset(&params->hd1_info, 0, sizeof(params->hd1_info));
134
135         /* Default sysdesc table */
136         params->sys_desc_table.length = 0;
137
138         if (image->type == KEXEC_TYPE_CRASH) {
139                 ret = crash_setup_memmap_entries(image, params);
140                 if (ret)
141                         return ret;
142         } else
143                 setup_memory_map_entries(params);
144
145         nr_e820_entries = params->e820_entries;
146
147         for (i = 0; i < nr_e820_entries; i++) {
148                 if (params->e820_map[i].type != E820_RAM)
149                         continue;
150                 start = params->e820_map[i].addr;
151                 end = params->e820_map[i].addr + params->e820_map[i].size - 1;
152
153                 if ((start <= 0x100000) && end > 0x100000) {
154                         mem_k = (end >> 10) - (0x100000 >> 10);
155                         params->screen_info.ext_mem_k = mem_k;
156                         params->alt_mem_k = mem_k;
157                         if (mem_k > 0xfc00)
158                                 params->screen_info.ext_mem_k = 0xfc00; /* 64M*/
159                         if (mem_k > 0xffffffff)
160                                 params->alt_mem_k = 0xffffffff;
161                 }
162         }
163
164         /* Setup EDD info */
165         memcpy(params->eddbuf, boot_params.eddbuf,
166                                 EDDMAXNR * sizeof(struct edd_info));
167         params->eddbuf_entries = boot_params.eddbuf_entries;
168
169         memcpy(params->edd_mbr_sig_buffer, boot_params.edd_mbr_sig_buffer,
170                EDD_MBR_SIG_MAX * sizeof(unsigned int));
171
172         return ret;
173 }
174
175 int bzImage64_probe(const char *buf, unsigned long len)
176 {
177         int ret = -ENOEXEC;
178         struct setup_header *header;
179
180         /* kernel should be atleast two sectors long */
181         if (len < 2 * 512) {
182                 pr_err("File is too short to be a bzImage\n");
183                 return ret;
184         }
185
186         header = (struct setup_header *)(buf + offsetof(struct boot_params, hdr));
187         if (memcmp((char *)&header->header, "HdrS", 4) != 0) {
188                 pr_err("Not a bzImage\n");
189                 return ret;
190         }
191
192         if (header->boot_flag != 0xAA55) {
193                 pr_err("No x86 boot sector present\n");
194                 return ret;
195         }
196
197         if (header->version < 0x020C) {
198                 pr_err("Must be at least protocol version 2.12\n");
199                 return ret;
200         }
201
202         if (!(header->loadflags & LOADED_HIGH)) {
203                 pr_err("zImage not a bzImage\n");
204                 return ret;
205         }
206
207         if (!(header->xloadflags & XLF_KERNEL_64)) {
208                 pr_err("Not a bzImage64. XLF_KERNEL_64 is not set.\n");
209                 return ret;
210         }
211
212         if (!(header->xloadflags & XLF_CAN_BE_LOADED_ABOVE_4G)) {
213                 pr_err("XLF_CAN_BE_LOADED_ABOVE_4G is not set.\n");
214                 return ret;
215         }
216
217         /* I've got a bzImage */
218         pr_debug("It's a relocatable bzImage64\n");
219         ret = 0;
220
221         return ret;
222 }
223
224 void *bzImage64_load(struct kimage *image, char *kernel,
225                      unsigned long kernel_len, char *initrd,
226                      unsigned long initrd_len, char *cmdline,
227                      unsigned long cmdline_len)
228 {
229
230         struct setup_header *header;
231         int setup_sects, kern16_size, ret = 0;
232         unsigned long setup_header_size, params_cmdline_sz;
233         struct boot_params *params;
234         unsigned long bootparam_load_addr, kernel_load_addr, initrd_load_addr;
235         unsigned long purgatory_load_addr;
236         unsigned long kernel_bufsz, kernel_memsz, kernel_align;
237         char *kernel_buf;
238         struct bzimage64_data *ldata;
239         struct kexec_entry64_regs regs64;
240         void *stack;
241         unsigned int setup_hdr_offset = offsetof(struct boot_params, hdr);
242
243         header = (struct setup_header *)(kernel + setup_hdr_offset);
244         setup_sects = header->setup_sects;
245         if (setup_sects == 0)
246                 setup_sects = 4;
247
248         kern16_size = (setup_sects + 1) * 512;
249         if (kernel_len < kern16_size) {
250                 pr_err("bzImage truncated\n");
251                 return ERR_PTR(-ENOEXEC);
252         }
253
254         if (cmdline_len > header->cmdline_size) {
255                 pr_err("Kernel command line too long\n");
256                 return ERR_PTR(-EINVAL);
257         }
258
259         /*
260          * In case of crash dump, we will append elfcorehdr=<addr> to
261          * command line. Make sure it does not overflow
262          */
263         if (cmdline_len + MAX_ELFCOREHDR_STR_LEN > header->cmdline_size) {
264                 pr_debug("Appending elfcorehdr=<addr> to command line exceeds maximum allowed length\n");
265                 return ERR_PTR(-EINVAL);
266         }
267
268         /* Allocate and load backup region */
269         if (image->type == KEXEC_TYPE_CRASH) {
270                 ret = crash_load_segments(image);
271                 if (ret)
272                         return ERR_PTR(ret);
273         }
274
275         /*
276          * Load purgatory. For 64bit entry point, purgatory  code can be
277          * anywhere.
278          */
279         ret = kexec_load_purgatory(image, MIN_PURGATORY_ADDR, ULONG_MAX, 1,
280                                    &purgatory_load_addr);
281         if (ret) {
282                 pr_err("Loading purgatory failed\n");
283                 return ERR_PTR(ret);
284         }
285
286         pr_debug("Loaded purgatory at 0x%lx\n", purgatory_load_addr);
287
288         /* Load Bootparams and cmdline */
289         params_cmdline_sz = sizeof(struct boot_params) + cmdline_len +
290                                 MAX_ELFCOREHDR_STR_LEN;
291         params = kzalloc(params_cmdline_sz, GFP_KERNEL);
292         if (!params)
293                 return ERR_PTR(-ENOMEM);
294
295         /* Copy setup header onto bootparams. Documentation/x86/boot.txt */
296         setup_header_size = 0x0202 + kernel[0x0201] - setup_hdr_offset;
297
298         /* Is there a limit on setup header size? */
299         memcpy(&params->hdr, (kernel + setup_hdr_offset), setup_header_size);
300
301         ret = kexec_add_buffer(image, (char *)params, params_cmdline_sz,
302                                params_cmdline_sz, 16, MIN_BOOTPARAM_ADDR,
303                                ULONG_MAX, 1, &bootparam_load_addr);
304         if (ret)
305                 goto out_free_params;
306         pr_debug("Loaded boot_param and command line at 0x%lx bufsz=0x%lx memsz=0x%lx\n",
307                  bootparam_load_addr, params_cmdline_sz, params_cmdline_sz);
308
309         /* Load kernel */
310         kernel_buf = kernel + kern16_size;
311         kernel_bufsz =  kernel_len - kern16_size;
312         kernel_memsz = PAGE_ALIGN(header->init_size);
313         kernel_align = header->kernel_alignment;
314
315         ret = kexec_add_buffer(image, kernel_buf,
316                                kernel_bufsz, kernel_memsz, kernel_align,
317                                MIN_KERNEL_LOAD_ADDR, ULONG_MAX, 1,
318                                &kernel_load_addr);
319         if (ret)
320                 goto out_free_params;
321
322         pr_debug("Loaded 64bit kernel at 0x%lx bufsz=0x%lx memsz=0x%lx\n",
323                  kernel_load_addr, kernel_memsz, kernel_memsz);
324
325         /* Load initrd high */
326         if (initrd) {
327                 ret = kexec_add_buffer(image, initrd, initrd_len, initrd_len,
328                                        PAGE_SIZE, MIN_INITRD_LOAD_ADDR,
329                                        ULONG_MAX, 1, &initrd_load_addr);
330                 if (ret)
331                         goto out_free_params;
332
333                 pr_debug("Loaded initrd at 0x%lx bufsz=0x%lx memsz=0x%lx\n",
334                                 initrd_load_addr, initrd_len, initrd_len);
335
336                 setup_initrd(params, initrd_load_addr, initrd_len);
337         }
338
339         setup_cmdline(image, params, bootparam_load_addr,
340                       sizeof(struct boot_params), cmdline, cmdline_len);
341
342         /* bootloader info. Do we need a separate ID for kexec kernel loader? */
343         params->hdr.type_of_loader = 0x0D << 4;
344         params->hdr.loadflags = 0;
345
346         /* Setup purgatory regs for entry */
347         ret = kexec_purgatory_get_set_symbol(image, "entry64_regs", &regs64,
348                                              sizeof(regs64), 1);
349         if (ret)
350                 goto out_free_params;
351
352         regs64.rbx = 0; /* Bootstrap Processor */
353         regs64.rsi = bootparam_load_addr;
354         regs64.rip = kernel_load_addr + 0x200;
355         stack = kexec_purgatory_get_symbol_addr(image, "stack_end");
356         if (IS_ERR(stack)) {
357                 pr_err("Could not find address of symbol stack_end\n");
358                 ret = -EINVAL;
359                 goto out_free_params;
360         }
361
362         regs64.rsp = (unsigned long)stack;
363         ret = kexec_purgatory_get_set_symbol(image, "entry64_regs", &regs64,
364                                              sizeof(regs64), 0);
365         if (ret)
366                 goto out_free_params;
367
368         ret = setup_boot_parameters(image, params);
369         if (ret)
370                 goto out_free_params;
371
372         /* Allocate loader specific data */
373         ldata = kzalloc(sizeof(struct bzimage64_data), GFP_KERNEL);
374         if (!ldata) {
375                 ret = -ENOMEM;
376                 goto out_free_params;
377         }
378
379         /*
380          * Store pointer to params so that it could be freed after loading
381          * params segment has been loaded and contents have been copied
382          * somewhere else.
383          */
384         ldata->bootparams_buf = params;
385         return ldata;
386
387 out_free_params:
388         kfree(params);
389         return ERR_PTR(ret);
390 }
391
392 /* This cleanup function is called after various segments have been loaded */
393 int bzImage64_cleanup(void *loader_data)
394 {
395         struct bzimage64_data *ldata = loader_data;
396
397         if (!ldata)
398                 return 0;
399
400         kfree(ldata->bootparams_buf);
401         ldata->bootparams_buf = NULL;
402
403         return 0;
404 }
405
406 struct kexec_file_ops kexec_bzImage64_ops = {
407         .probe = bzImage64_probe,
408         .load = bzImage64_load,
409         .cleanup = bzImage64_cleanup,
410 };