• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 #define pr_fmt(fmt) "efi: " fmt
2 
3 #include <linux/init.h>
4 #include <linux/kernel.h>
5 #include <linux/string.h>
6 #include <linux/time.h>
7 #include <linux/types.h>
8 #include <linux/efi.h>
9 #include <linux/slab.h>
10 #include <linux/memblock.h>
11 #include <linux/bootmem.h>
12 #include <linux/acpi.h>
13 #include <linux/dmi.h>
14 #include <asm/efi.h>
15 #include <asm/uv/uv.h>
16 
17 #define EFI_MIN_RESERVE 5120
18 
19 #define EFI_DUMMY_GUID \
20 	EFI_GUID(0x4424ac57, 0xbe4b, 0x47dd, 0x9e, 0x97, 0xed, 0x50, 0xf0, 0x9f, 0x92, 0xa9)
21 
22 static efi_char16_t efi_dummy_name[6] = { 'D', 'U', 'M', 'M', 'Y', 0 };
23 
24 static bool efi_no_storage_paranoia;
25 
26 /*
27  * Some firmware implementations refuse to boot if there's insufficient
28  * space in the variable store. The implementation of garbage collection
29  * in some FW versions causes stale (deleted) variables to take up space
30  * longer than intended and space is only freed once the store becomes
31  * almost completely full.
32  *
33  * Enabling this option disables the space checks in
34  * efi_query_variable_store() and forces garbage collection.
35  *
36  * Only enable this option if deleting EFI variables does not free up
37  * space in your variable store, e.g. if despite deleting variables
38  * you're unable to create new ones.
39  */
setup_storage_paranoia(char * arg)40 static int __init setup_storage_paranoia(char *arg)
41 {
42 	efi_no_storage_paranoia = true;
43 	return 0;
44 }
45 early_param("efi_no_storage_paranoia", setup_storage_paranoia);
46 
47 /*
48  * Deleting the dummy variable which kicks off garbage collection
49 */
efi_delete_dummy_variable(void)50 void efi_delete_dummy_variable(void)
51 {
52 	efi.set_variable(efi_dummy_name, &EFI_DUMMY_GUID,
53 			 EFI_VARIABLE_NON_VOLATILE |
54 			 EFI_VARIABLE_BOOTSERVICE_ACCESS |
55 			 EFI_VARIABLE_RUNTIME_ACCESS,
56 			 0, NULL);
57 }
58 
59 /*
60  * In the nonblocking case we do not attempt to perform garbage
61  * collection if we do not have enough free space. Rather, we do the
62  * bare minimum check and give up immediately if the available space
63  * is below EFI_MIN_RESERVE.
64  *
65  * This function is intended to be small and simple because it is
66  * invoked from crash handler paths.
67  */
68 static efi_status_t
query_variable_store_nonblocking(u32 attributes,unsigned long size)69 query_variable_store_nonblocking(u32 attributes, unsigned long size)
70 {
71 	efi_status_t status;
72 	u64 storage_size, remaining_size, max_size;
73 
74 	status = efi.query_variable_info_nonblocking(attributes, &storage_size,
75 						     &remaining_size,
76 						     &max_size);
77 	if (status != EFI_SUCCESS)
78 		return status;
79 
80 	if (remaining_size - size < EFI_MIN_RESERVE)
81 		return EFI_OUT_OF_RESOURCES;
82 
83 	return EFI_SUCCESS;
84 }
85 
86 /*
87  * Some firmware implementations refuse to boot if there's insufficient space
88  * in the variable store. Ensure that we never use more than a safe limit.
89  *
90  * Return EFI_SUCCESS if it is safe to write 'size' bytes to the variable
91  * store.
92  */
efi_query_variable_store(u32 attributes,unsigned long size,bool nonblocking)93 efi_status_t efi_query_variable_store(u32 attributes, unsigned long size,
94 				      bool nonblocking)
95 {
96 	efi_status_t status;
97 	u64 storage_size, remaining_size, max_size;
98 
99 	if (!(attributes & EFI_VARIABLE_NON_VOLATILE))
100 		return 0;
101 
102 	if (nonblocking)
103 		return query_variable_store_nonblocking(attributes, size);
104 
105 	status = efi.query_variable_info(attributes, &storage_size,
106 					 &remaining_size, &max_size);
107 	if (status != EFI_SUCCESS)
108 		return status;
109 
110 	/*
111 	 * We account for that by refusing the write if permitting it would
112 	 * reduce the available space to under 5KB. This figure was provided by
113 	 * Samsung, so should be safe.
114 	 */
115 	if ((remaining_size - size < EFI_MIN_RESERVE) &&
116 		!efi_no_storage_paranoia) {
117 
118 		/*
119 		 * Triggering garbage collection may require that the firmware
120 		 * generate a real EFI_OUT_OF_RESOURCES error. We can force
121 		 * that by attempting to use more space than is available.
122 		 */
123 		unsigned long dummy_size = remaining_size + 1024;
124 		void *dummy = kzalloc(dummy_size, GFP_ATOMIC);
125 
126 		if (!dummy)
127 			return EFI_OUT_OF_RESOURCES;
128 
129 		status = efi.set_variable(efi_dummy_name, &EFI_DUMMY_GUID,
130 					  EFI_VARIABLE_NON_VOLATILE |
131 					  EFI_VARIABLE_BOOTSERVICE_ACCESS |
132 					  EFI_VARIABLE_RUNTIME_ACCESS,
133 					  dummy_size, dummy);
134 
135 		if (status == EFI_SUCCESS) {
136 			/*
137 			 * This should have failed, so if it didn't make sure
138 			 * that we delete it...
139 			 */
140 			efi_delete_dummy_variable();
141 		}
142 
143 		kfree(dummy);
144 
145 		/*
146 		 * The runtime code may now have triggered a garbage collection
147 		 * run, so check the variable info again
148 		 */
149 		status = efi.query_variable_info(attributes, &storage_size,
150 						 &remaining_size, &max_size);
151 
152 		if (status != EFI_SUCCESS)
153 			return status;
154 
155 		/*
156 		 * There still isn't enough room, so return an error
157 		 */
158 		if (remaining_size - size < EFI_MIN_RESERVE)
159 			return EFI_OUT_OF_RESOURCES;
160 	}
161 
162 	return EFI_SUCCESS;
163 }
164 EXPORT_SYMBOL_GPL(efi_query_variable_store);
165 
166 /*
167  * The UEFI specification makes it clear that the operating system is
168  * free to do whatever it wants with boot services code after
169  * ExitBootServices() has been called. Ignoring this recommendation a
170  * significant bunch of EFI implementations continue calling into boot
171  * services code (SetVirtualAddressMap). In order to work around such
172  * buggy implementations we reserve boot services region during EFI
173  * init and make sure it stays executable. Then, after
174  * SetVirtualAddressMap(), it is discarded.
175  *
176  * However, some boot services regions contain data that is required
177  * by drivers, so we need to track which memory ranges can never be
178  * freed. This is done by tagging those regions with the
179  * EFI_MEMORY_RUNTIME attribute.
180  *
181  * Any driver that wants to mark a region as reserved must use
182  * efi_mem_reserve() which will insert a new EFI memory descriptor
183  * into efi.memmap (splitting existing regions if necessary) and tag
184  * it with EFI_MEMORY_RUNTIME.
185  */
efi_arch_mem_reserve(phys_addr_t addr,u64 size)186 void __init efi_arch_mem_reserve(phys_addr_t addr, u64 size)
187 {
188 	phys_addr_t new_phys, new_size;
189 	struct efi_mem_range mr;
190 	efi_memory_desc_t md;
191 	int num_entries;
192 	void *new;
193 
194 	if (efi_mem_desc_lookup(addr, &md)) {
195 		pr_err("Failed to lookup EFI memory descriptor for %pa\n", &addr);
196 		return;
197 	}
198 
199 	if (addr + size > md.phys_addr + (md.num_pages << EFI_PAGE_SHIFT)) {
200 		pr_err("Region spans EFI memory descriptors, %pa\n", &addr);
201 		return;
202 	}
203 
204 	/* No need to reserve regions that will never be freed. */
205 	if (md.attribute & EFI_MEMORY_RUNTIME)
206 		return;
207 
208 	size += addr % EFI_PAGE_SIZE;
209 	size = round_up(size, EFI_PAGE_SIZE);
210 	addr = round_down(addr, EFI_PAGE_SIZE);
211 
212 	mr.range.start = addr;
213 	mr.range.end = addr + size - 1;
214 	mr.attribute = md.attribute | EFI_MEMORY_RUNTIME;
215 
216 	num_entries = efi_memmap_split_count(&md, &mr.range);
217 	num_entries += efi.memmap.nr_map;
218 
219 	new_size = efi.memmap.desc_size * num_entries;
220 
221 	new_phys = efi_memmap_alloc(num_entries);
222 	if (!new_phys) {
223 		pr_err("Could not allocate boot services memmap\n");
224 		return;
225 	}
226 
227 	new = early_memremap(new_phys, new_size);
228 	if (!new) {
229 		pr_err("Failed to map new boot services memmap\n");
230 		return;
231 	}
232 
233 	efi_memmap_insert(&efi.memmap, new, &mr);
234 	early_memunmap(new, new_size);
235 
236 	efi_memmap_install(new_phys, num_entries);
237 }
238 
239 /*
240  * Helper function for efi_reserve_boot_services() to figure out if we
241  * can free regions in efi_free_boot_services().
242  *
243  * Use this function to ensure we do not free regions owned by somebody
244  * else. We must only reserve (and then free) regions:
245  *
246  * - Not within any part of the kernel
247  * - Not the BIOS reserved area (E820_RESERVED, E820_NVS, etc)
248  */
can_free_region(u64 start,u64 size)249 static bool can_free_region(u64 start, u64 size)
250 {
251 	if (start + size > __pa_symbol(_text) && start <= __pa_symbol(_end))
252 		return false;
253 
254 	if (!e820_all_mapped(start, start+size, E820_RAM))
255 		return false;
256 
257 	return true;
258 }
259 
efi_reserve_boot_services(void)260 void __init efi_reserve_boot_services(void)
261 {
262 	efi_memory_desc_t *md;
263 
264 	for_each_efi_memory_desc(md) {
265 		u64 start = md->phys_addr;
266 		u64 size = md->num_pages << EFI_PAGE_SHIFT;
267 		bool already_reserved;
268 
269 		if (md->type != EFI_BOOT_SERVICES_CODE &&
270 		    md->type != EFI_BOOT_SERVICES_DATA)
271 			continue;
272 
273 		already_reserved = memblock_is_region_reserved(start, size);
274 
275 		/*
276 		 * Because the following memblock_reserve() is paired
277 		 * with free_bootmem_late() for this region in
278 		 * efi_free_boot_services(), we must be extremely
279 		 * careful not to reserve, and subsequently free,
280 		 * critical regions of memory (like the kernel image) or
281 		 * those regions that somebody else has already
282 		 * reserved.
283 		 *
284 		 * A good example of a critical region that must not be
285 		 * freed is page zero (first 4Kb of memory), which may
286 		 * contain boot services code/data but is marked
287 		 * E820_RESERVED by trim_bios_range().
288 		 */
289 		if (!already_reserved) {
290 			memblock_reserve(start, size);
291 
292 			/*
293 			 * If we are the first to reserve the region, no
294 			 * one else cares about it. We own it and can
295 			 * free it later.
296 			 */
297 			if (can_free_region(start, size))
298 				continue;
299 		}
300 
301 		/*
302 		 * We don't own the region. We must not free it.
303 		 *
304 		 * Setting this bit for a boot services region really
305 		 * doesn't make sense as far as the firmware is
306 		 * concerned, but it does provide us with a way to tag
307 		 * those regions that must not be paired with
308 		 * free_bootmem_late().
309 		 */
310 		md->attribute |= EFI_MEMORY_RUNTIME;
311 	}
312 }
313 
efi_free_boot_services(void)314 void __init efi_free_boot_services(void)
315 {
316 	phys_addr_t new_phys, new_size;
317 	efi_memory_desc_t *md;
318 	int num_entries = 0;
319 	void *new, *new_md;
320 
321 	for_each_efi_memory_desc(md) {
322 		unsigned long long start = md->phys_addr;
323 		unsigned long long size = md->num_pages << EFI_PAGE_SHIFT;
324 		size_t rm_size;
325 
326 		if (md->type != EFI_BOOT_SERVICES_CODE &&
327 		    md->type != EFI_BOOT_SERVICES_DATA) {
328 			num_entries++;
329 			continue;
330 		}
331 
332 		/* Do not free, someone else owns it: */
333 		if (md->attribute & EFI_MEMORY_RUNTIME) {
334 			num_entries++;
335 			continue;
336 		}
337 
338 		/*
339 		 * Nasty quirk: if all sub-1MB memory is used for boot
340 		 * services, we can get here without having allocated the
341 		 * real mode trampoline.  It's too late to hand boot services
342 		 * memory back to the memblock allocator, so instead
343 		 * try to manually allocate the trampoline if needed.
344 		 *
345 		 * I've seen this on a Dell XPS 13 9350 with firmware
346 		 * 1.4.4 with SGX enabled booting Linux via Fedora 24's
347 		 * grub2-efi on a hard disk.  (And no, I don't know why
348 		 * this happened, but Linux should still try to boot rather
349 		 * panicing early.)
350 		 */
351 		rm_size = real_mode_size_needed();
352 		if (rm_size && (start + rm_size) < (1<<20) && size >= rm_size) {
353 			set_real_mode_mem(start, rm_size);
354 			start += rm_size;
355 			size -= rm_size;
356 		}
357 
358 		free_bootmem_late(start, size);
359 	}
360 
361 	if (!num_entries)
362 		return;
363 
364 	new_size = efi.memmap.desc_size * num_entries;
365 	new_phys = efi_memmap_alloc(num_entries);
366 	if (!new_phys) {
367 		pr_err("Failed to allocate new EFI memmap\n");
368 		return;
369 	}
370 
371 	new = memremap(new_phys, new_size, MEMREMAP_WB);
372 	if (!new) {
373 		pr_err("Failed to map new EFI memmap\n");
374 		return;
375 	}
376 
377 	/*
378 	 * Build a new EFI memmap that excludes any boot services
379 	 * regions that are not tagged EFI_MEMORY_RUNTIME, since those
380 	 * regions have now been freed.
381 	 */
382 	new_md = new;
383 	for_each_efi_memory_desc(md) {
384 		if (!(md->attribute & EFI_MEMORY_RUNTIME) &&
385 		    (md->type == EFI_BOOT_SERVICES_CODE ||
386 		     md->type == EFI_BOOT_SERVICES_DATA))
387 			continue;
388 
389 		memcpy(new_md, md, efi.memmap.desc_size);
390 		new_md += efi.memmap.desc_size;
391 	}
392 
393 	memunmap(new);
394 
395 	if (efi_memmap_install(new_phys, num_entries)) {
396 		pr_err("Could not install new EFI memmap\n");
397 		return;
398 	}
399 }
400 
401 /*
402  * A number of config table entries get remapped to virtual addresses
403  * after entering EFI virtual mode. However, the kexec kernel requires
404  * their physical addresses therefore we pass them via setup_data and
405  * correct those entries to their respective physical addresses here.
406  *
407  * Currently only handles smbios which is necessary for some firmware
408  * implementation.
409  */
efi_reuse_config(u64 tables,int nr_tables)410 int __init efi_reuse_config(u64 tables, int nr_tables)
411 {
412 	int i, sz, ret = 0;
413 	void *p, *tablep;
414 	struct efi_setup_data *data;
415 
416 	if (!efi_setup)
417 		return 0;
418 
419 	if (!efi_enabled(EFI_64BIT))
420 		return 0;
421 
422 	data = early_memremap(efi_setup, sizeof(*data));
423 	if (!data) {
424 		ret = -ENOMEM;
425 		goto out;
426 	}
427 
428 	if (!data->smbios)
429 		goto out_memremap;
430 
431 	sz = sizeof(efi_config_table_64_t);
432 
433 	p = tablep = early_memremap(tables, nr_tables * sz);
434 	if (!p) {
435 		pr_err("Could not map Configuration table!\n");
436 		ret = -ENOMEM;
437 		goto out_memremap;
438 	}
439 
440 	for (i = 0; i < efi.systab->nr_tables; i++) {
441 		efi_guid_t guid;
442 
443 		guid = ((efi_config_table_64_t *)p)->guid;
444 
445 		if (!efi_guidcmp(guid, SMBIOS_TABLE_GUID))
446 			((efi_config_table_64_t *)p)->table = data->smbios;
447 		p += sz;
448 	}
449 	early_memunmap(tablep, nr_tables * sz);
450 
451 out_memremap:
452 	early_memunmap(data, sizeof(*data));
453 out:
454 	return ret;
455 }
456 
457 static const struct dmi_system_id sgi_uv1_dmi[] = {
458 	{ NULL, "SGI UV1",
459 		{	DMI_MATCH(DMI_PRODUCT_NAME,	"Stoutland Platform"),
460 			DMI_MATCH(DMI_PRODUCT_VERSION,	"1.0"),
461 			DMI_MATCH(DMI_BIOS_VENDOR,	"SGI.COM"),
462 		}
463 	},
464 	{ } /* NULL entry stops DMI scanning */
465 };
466 
efi_apply_memmap_quirks(void)467 void __init efi_apply_memmap_quirks(void)
468 {
469 	/*
470 	 * Once setup is done earlier, unmap the EFI memory map on mismatched
471 	 * firmware/kernel architectures since there is no support for runtime
472 	 * services.
473 	 */
474 	if (!efi_runtime_supported()) {
475 		pr_info("Setup done, disabling due to 32/64-bit mismatch\n");
476 		efi_memmap_unmap();
477 	}
478 
479 	/* UV2+ BIOS has a fix for this issue.  UV1 still needs the quirk. */
480 	if (dmi_check_system(sgi_uv1_dmi))
481 		set_bit(EFI_OLD_MEMMAP, &efi.flags);
482 }
483 
484 /*
485  * For most modern platforms the preferred method of powering off is via
486  * ACPI. However, there are some that are known to require the use of
487  * EFI runtime services and for which ACPI does not work at all.
488  *
489  * Using EFI is a last resort, to be used only if no other option
490  * exists.
491  */
efi_reboot_required(void)492 bool efi_reboot_required(void)
493 {
494 	if (!acpi_gbl_reduced_hardware)
495 		return false;
496 
497 	efi_reboot_quirk_mode = EFI_RESET_WARM;
498 	return true;
499 }
500 
efi_poweroff_required(void)501 bool efi_poweroff_required(void)
502 {
503 	return acpi_gbl_reduced_hardware || acpi_no_s5;
504 }
505