1 // SPDX-License-Identifier: GPL-2.0-only
2 /*
3 * AMD CPU Microcode Update Driver for Linux
4 *
5 * This driver allows to upgrade microcode on F10h AMD
6 * CPUs and later.
7 *
8 * Copyright (C) 2008-2011 Advanced Micro Devices Inc.
9 * 2013-2018 Borislav Petkov <bp@alien8.de>
10 *
11 * Author: Peter Oruba <peter.oruba@amd.com>
12 *
13 * Based on work by:
14 * Tigran Aivazian <aivazian.tigran@gmail.com>
15 *
16 * early loader:
17 * Copyright (C) 2013 Advanced Micro Devices, Inc.
18 *
19 * Author: Jacob Shin <jacob.shin@amd.com>
20 * Fixes: Borislav Petkov <bp@suse.de>
21 */
22 #define pr_fmt(fmt) "microcode: " fmt
23
24 #include <linux/earlycpio.h>
25 #include <linux/firmware.h>
26 #include <linux/bsearch.h>
27 #include <linux/uaccess.h>
28 #include <linux/vmalloc.h>
29 #include <linux/initrd.h>
30 #include <linux/kernel.h>
31 #include <linux/pci.h>
32
33 #include <crypto/sha2.h>
34
35 #include <asm/microcode.h>
36 #include <asm/processor.h>
37 #include <asm/cmdline.h>
38 #include <asm/setup.h>
39 #include <asm/cpu.h>
40 #include <asm/msr.h>
41 #include <asm/tlb.h>
42
43 #include "internal.h"
44
45 struct ucode_patch {
46 struct list_head plist;
47 void *data;
48 unsigned int size;
49 u32 patch_id;
50 u16 equiv_cpu;
51 };
52
53 static LIST_HEAD(microcode_cache);
54
55 #define UCODE_MAGIC 0x00414d44
56 #define UCODE_EQUIV_CPU_TABLE_TYPE 0x00000000
57 #define UCODE_UCODE_TYPE 0x00000001
58
59 #define SECTION_HDR_SIZE 8
60 #define CONTAINER_HDR_SZ 12
61
62 struct equiv_cpu_entry {
63 u32 installed_cpu;
64 u32 fixed_errata_mask;
65 u32 fixed_errata_compare;
66 u16 equiv_cpu;
67 u16 res;
68 } __packed;
69
70 struct microcode_header_amd {
71 u32 data_code;
72 u32 patch_id;
73 u16 mc_patch_data_id;
74 u8 mc_patch_data_len;
75 u8 init_flag;
76 u32 mc_patch_data_checksum;
77 u32 nb_dev_id;
78 u32 sb_dev_id;
79 u16 processor_rev_id;
80 u8 nb_rev_id;
81 u8 sb_rev_id;
82 u8 bios_api_rev;
83 u8 reserved1[3];
84 u32 match_reg[8];
85 } __packed;
86
87 struct microcode_amd {
88 struct microcode_header_amd hdr;
89 unsigned int mpb[];
90 };
91
92 static struct equiv_cpu_table {
93 unsigned int num_entries;
94 struct equiv_cpu_entry *entry;
95 } equiv_table;
96
97 union zen_patch_rev {
98 struct {
99 __u32 rev : 8,
100 stepping : 4,
101 model : 4,
102 __reserved : 4,
103 ext_model : 4,
104 ext_fam : 8;
105 };
106 __u32 ucode_rev;
107 };
108
109 union cpuid_1_eax {
110 struct {
111 __u32 stepping : 4,
112 model : 4,
113 family : 4,
114 __reserved0 : 4,
115 ext_model : 4,
116 ext_fam : 8,
117 __reserved1 : 4;
118 };
119 __u32 full;
120 };
121
122 /*
123 * This points to the current valid container of microcode patches which we will
124 * save from the initrd/builtin before jettisoning its contents. @mc is the
125 * microcode patch we found to match.
126 */
127 struct cont_desc {
128 struct microcode_amd *mc;
129 u32 psize;
130 u8 *data;
131 size_t size;
132 };
133
134 /*
135 * Microcode patch container file is prepended to the initrd in cpio
136 * format. See Documentation/arch/x86/microcode.rst
137 */
138 static const char
139 ucode_path[] __maybe_unused = "kernel/x86/microcode/AuthenticAMD.bin";
140
141 /*
142 * This is CPUID(1).EAX on the BSP. It is used in two ways:
143 *
144 * 1. To ignore the equivalence table on Zen1 and newer.
145 *
146 * 2. To match which patches to load because the patch revision ID
147 * already contains the f/m/s for which the microcode is destined
148 * for.
149 */
150 static u32 bsp_cpuid_1_eax __ro_after_init;
151
152 static bool sha_check = true;
153
154 struct patch_digest {
155 u32 patch_id;
156 u8 sha256[SHA256_DIGEST_SIZE];
157 };
158
159 #include "amd_shas.c"
160
cmp_id(const void * key,const void * elem)161 static int cmp_id(const void *key, const void *elem)
162 {
163 struct patch_digest *pd = (struct patch_digest *)elem;
164 u32 patch_id = *(u32 *)key;
165
166 if (patch_id == pd->patch_id)
167 return 0;
168 else if (patch_id < pd->patch_id)
169 return -1;
170 else
171 return 1;
172 }
173
cpuid_to_ucode_rev(unsigned int val)174 static u32 cpuid_to_ucode_rev(unsigned int val)
175 {
176 union zen_patch_rev p = {};
177 union cpuid_1_eax c;
178
179 c.full = val;
180
181 p.stepping = c.stepping;
182 p.model = c.model;
183 p.ext_model = c.ext_model;
184 p.ext_fam = c.ext_fam;
185
186 return p.ucode_rev;
187 }
188
need_sha_check(u32 cur_rev)189 static bool need_sha_check(u32 cur_rev)
190 {
191 if (!cur_rev) {
192 cur_rev = cpuid_to_ucode_rev(bsp_cpuid_1_eax);
193 pr_info_once("No current revision, generating the lowest one: 0x%x\n", cur_rev);
194 }
195
196 switch (cur_rev >> 8) {
197 case 0x80012: return cur_rev <= 0x800126f; break;
198 case 0x80082: return cur_rev <= 0x800820f; break;
199 case 0x83010: return cur_rev <= 0x830107c; break;
200 case 0x86001: return cur_rev <= 0x860010e; break;
201 case 0x86081: return cur_rev <= 0x8608108; break;
202 case 0x87010: return cur_rev <= 0x8701034; break;
203 case 0x8a000: return cur_rev <= 0x8a0000a; break;
204 case 0xa0010: return cur_rev <= 0xa00107a; break;
205 case 0xa0011: return cur_rev <= 0xa0011da; break;
206 case 0xa0012: return cur_rev <= 0xa001243; break;
207 case 0xa0082: return cur_rev <= 0xa00820e; break;
208 case 0xa1011: return cur_rev <= 0xa101153; break;
209 case 0xa1012: return cur_rev <= 0xa10124e; break;
210 case 0xa1081: return cur_rev <= 0xa108109; break;
211 case 0xa2010: return cur_rev <= 0xa20102f; break;
212 case 0xa2012: return cur_rev <= 0xa201212; break;
213 case 0xa4041: return cur_rev <= 0xa404109; break;
214 case 0xa5000: return cur_rev <= 0xa500013; break;
215 case 0xa6012: return cur_rev <= 0xa60120a; break;
216 case 0xa7041: return cur_rev <= 0xa704109; break;
217 case 0xa7052: return cur_rev <= 0xa705208; break;
218 case 0xa7080: return cur_rev <= 0xa708009; break;
219 case 0xa70c0: return cur_rev <= 0xa70C009; break;
220 case 0xaa001: return cur_rev <= 0xaa00116; break;
221 case 0xaa002: return cur_rev <= 0xaa00218; break;
222 case 0xb0021: return cur_rev <= 0xb002146; break;
223 case 0xb1010: return cur_rev <= 0xb101046; break;
224 case 0xb2040: return cur_rev <= 0xb204031; break;
225 case 0xb4040: return cur_rev <= 0xb404031; break;
226 case 0xb6000: return cur_rev <= 0xb600031; break;
227 case 0xb7000: return cur_rev <= 0xb700031; break;
228 default: break;
229 }
230
231 pr_info("You should not be seeing this. Please send the following couple of lines to x86-<at>-kernel.org\n");
232 pr_info("CPUID(1).EAX: 0x%x, current revision: 0x%x\n", bsp_cpuid_1_eax, cur_rev);
233 return true;
234 }
235
verify_sha256_digest(u32 patch_id,u32 cur_rev,const u8 * data,unsigned int len)236 static bool verify_sha256_digest(u32 patch_id, u32 cur_rev, const u8 *data, unsigned int len)
237 {
238 struct patch_digest *pd = NULL;
239 u8 digest[SHA256_DIGEST_SIZE];
240 int i;
241
242 if (x86_family(bsp_cpuid_1_eax) < 0x17)
243 return true;
244
245 if (!need_sha_check(cur_rev))
246 return true;
247
248 if (!sha_check)
249 return true;
250
251 pd = bsearch(&patch_id, phashes, ARRAY_SIZE(phashes), sizeof(struct patch_digest), cmp_id);
252 if (!pd) {
253 pr_err("No sha256 digest for patch ID: 0x%x found\n", patch_id);
254 return false;
255 }
256
257 sha256(data, len, digest);
258
259 if (memcmp(digest, pd->sha256, sizeof(digest))) {
260 pr_err("Patch 0x%x SHA256 digest mismatch!\n", patch_id);
261
262 for (i = 0; i < SHA256_DIGEST_SIZE; i++)
263 pr_cont("0x%x ", digest[i]);
264 pr_info("\n");
265
266 return false;
267 }
268
269 return true;
270 }
271
get_patch_level(void)272 static u32 get_patch_level(void)
273 {
274 u32 rev, dummy __always_unused;
275
276 native_rdmsr(MSR_AMD64_PATCH_LEVEL, rev, dummy);
277
278 return rev;
279 }
280
ucode_rev_to_cpuid(unsigned int val)281 static union cpuid_1_eax ucode_rev_to_cpuid(unsigned int val)
282 {
283 union zen_patch_rev p;
284 union cpuid_1_eax c;
285
286 p.ucode_rev = val;
287 c.full = 0;
288
289 c.stepping = p.stepping;
290 c.model = p.model;
291 c.ext_model = p.ext_model;
292 c.family = 0xf;
293 c.ext_fam = p.ext_fam;
294
295 return c;
296 }
297
find_equiv_id(struct equiv_cpu_table * et,u32 sig)298 static u16 find_equiv_id(struct equiv_cpu_table *et, u32 sig)
299 {
300 unsigned int i;
301
302 /* Zen and newer do not need an equivalence table. */
303 if (x86_family(bsp_cpuid_1_eax) >= 0x17)
304 return 0;
305
306 if (!et || !et->num_entries)
307 return 0;
308
309 for (i = 0; i < et->num_entries; i++) {
310 struct equiv_cpu_entry *e = &et->entry[i];
311
312 if (sig == e->installed_cpu)
313 return e->equiv_cpu;
314 }
315 return 0;
316 }
317
318 /*
319 * Check whether there is a valid microcode container file at the beginning
320 * of @buf of size @buf_size.
321 */
verify_container(const u8 * buf,size_t buf_size)322 static bool verify_container(const u8 *buf, size_t buf_size)
323 {
324 u32 cont_magic;
325
326 if (buf_size <= CONTAINER_HDR_SZ) {
327 pr_debug("Truncated microcode container header.\n");
328 return false;
329 }
330
331 cont_magic = *(const u32 *)buf;
332 if (cont_magic != UCODE_MAGIC) {
333 pr_debug("Invalid magic value (0x%08x).\n", cont_magic);
334 return false;
335 }
336
337 return true;
338 }
339
340 /*
341 * Check whether there is a valid, non-truncated CPU equivalence table at the
342 * beginning of @buf of size @buf_size.
343 */
verify_equivalence_table(const u8 * buf,size_t buf_size)344 static bool verify_equivalence_table(const u8 *buf, size_t buf_size)
345 {
346 const u32 *hdr = (const u32 *)buf;
347 u32 cont_type, equiv_tbl_len;
348
349 if (!verify_container(buf, buf_size))
350 return false;
351
352 /* Zen and newer do not need an equivalence table. */
353 if (x86_family(bsp_cpuid_1_eax) >= 0x17)
354 return true;
355
356 cont_type = hdr[1];
357 if (cont_type != UCODE_EQUIV_CPU_TABLE_TYPE) {
358 pr_debug("Wrong microcode container equivalence table type: %u.\n",
359 cont_type);
360 return false;
361 }
362
363 buf_size -= CONTAINER_HDR_SZ;
364
365 equiv_tbl_len = hdr[2];
366 if (equiv_tbl_len < sizeof(struct equiv_cpu_entry) ||
367 buf_size < equiv_tbl_len) {
368 pr_debug("Truncated equivalence table.\n");
369 return false;
370 }
371
372 return true;
373 }
374
375 /*
376 * Check whether there is a valid, non-truncated microcode patch section at the
377 * beginning of @buf of size @buf_size.
378 *
379 * On success, @sh_psize returns the patch size according to the section header,
380 * to the caller.
381 */
__verify_patch_section(const u8 * buf,size_t buf_size,u32 * sh_psize)382 static bool __verify_patch_section(const u8 *buf, size_t buf_size, u32 *sh_psize)
383 {
384 u32 p_type, p_size;
385 const u32 *hdr;
386
387 if (buf_size < SECTION_HDR_SIZE) {
388 pr_debug("Truncated patch section.\n");
389 return false;
390 }
391
392 hdr = (const u32 *)buf;
393 p_type = hdr[0];
394 p_size = hdr[1];
395
396 if (p_type != UCODE_UCODE_TYPE) {
397 pr_debug("Invalid type field (0x%x) in container file section header.\n",
398 p_type);
399 return false;
400 }
401
402 if (p_size < sizeof(struct microcode_header_amd)) {
403 pr_debug("Patch of size %u too short.\n", p_size);
404 return false;
405 }
406
407 *sh_psize = p_size;
408
409 return true;
410 }
411
412 /*
413 * Check whether the passed remaining file @buf_size is large enough to contain
414 * a patch of the indicated @sh_psize (and also whether this size does not
415 * exceed the per-family maximum). @sh_psize is the size read from the section
416 * header.
417 */
__verify_patch_size(u32 sh_psize,size_t buf_size)418 static bool __verify_patch_size(u32 sh_psize, size_t buf_size)
419 {
420 u8 family = x86_family(bsp_cpuid_1_eax);
421 u32 max_size;
422
423 if (family >= 0x15)
424 goto ret;
425
426 #define F1XH_MPB_MAX_SIZE 2048
427 #define F14H_MPB_MAX_SIZE 1824
428
429 switch (family) {
430 case 0x10 ... 0x12:
431 max_size = F1XH_MPB_MAX_SIZE;
432 break;
433 case 0x14:
434 max_size = F14H_MPB_MAX_SIZE;
435 break;
436 default:
437 WARN(1, "%s: WTF family: 0x%x\n", __func__, family);
438 return false;
439 }
440
441 if (sh_psize > max_size)
442 return false;
443
444 ret:
445 /* Working with the whole buffer so < is ok. */
446 return sh_psize <= buf_size;
447 }
448
449 /*
450 * Verify the patch in @buf.
451 *
452 * Returns:
453 * negative: on error
454 * positive: patch is not for this family, skip it
455 * 0: success
456 */
verify_patch(const u8 * buf,size_t buf_size,u32 * patch_size)457 static int verify_patch(const u8 *buf, size_t buf_size, u32 *patch_size)
458 {
459 u8 family = x86_family(bsp_cpuid_1_eax);
460 struct microcode_header_amd *mc_hdr;
461 u32 sh_psize;
462 u16 proc_id;
463 u8 patch_fam;
464
465 if (!__verify_patch_section(buf, buf_size, &sh_psize))
466 return -1;
467
468 /*
469 * The section header length is not included in this indicated size
470 * but is present in the leftover file length so we need to subtract
471 * it before passing this value to the function below.
472 */
473 buf_size -= SECTION_HDR_SIZE;
474
475 /*
476 * Check if the remaining buffer is big enough to contain a patch of
477 * size sh_psize, as the section claims.
478 */
479 if (buf_size < sh_psize) {
480 pr_debug("Patch of size %u truncated.\n", sh_psize);
481 return -1;
482 }
483
484 if (!__verify_patch_size(sh_psize, buf_size)) {
485 pr_debug("Per-family patch size mismatch.\n");
486 return -1;
487 }
488
489 *patch_size = sh_psize;
490
491 mc_hdr = (struct microcode_header_amd *)(buf + SECTION_HDR_SIZE);
492 if (mc_hdr->nb_dev_id || mc_hdr->sb_dev_id) {
493 pr_err("Patch-ID 0x%08x: chipset-specific code unsupported.\n", mc_hdr->patch_id);
494 return -1;
495 }
496
497 proc_id = mc_hdr->processor_rev_id;
498 patch_fam = 0xf + (proc_id >> 12);
499 if (patch_fam != family)
500 return 1;
501
502 return 0;
503 }
504
mc_patch_matches(struct microcode_amd * mc,u16 eq_id)505 static bool mc_patch_matches(struct microcode_amd *mc, u16 eq_id)
506 {
507 /* Zen and newer do not need an equivalence table. */
508 if (x86_family(bsp_cpuid_1_eax) >= 0x17)
509 return ucode_rev_to_cpuid(mc->hdr.patch_id).full == bsp_cpuid_1_eax;
510 else
511 return eq_id == mc->hdr.processor_rev_id;
512 }
513
514 /*
515 * This scans the ucode blob for the proper container as we can have multiple
516 * containers glued together.
517 *
518 * Returns the amount of bytes consumed while scanning. @desc contains all the
519 * data we're going to use in later stages of the application.
520 */
parse_container(u8 * ucode,size_t size,struct cont_desc * desc)521 static size_t parse_container(u8 *ucode, size_t size, struct cont_desc *desc)
522 {
523 struct equiv_cpu_table table;
524 size_t orig_size = size;
525 u32 *hdr = (u32 *)ucode;
526 u16 eq_id;
527 u8 *buf;
528
529 if (!verify_equivalence_table(ucode, size))
530 return 0;
531
532 buf = ucode;
533
534 table.entry = (struct equiv_cpu_entry *)(buf + CONTAINER_HDR_SZ);
535 table.num_entries = hdr[2] / sizeof(struct equiv_cpu_entry);
536
537 /*
538 * Find the equivalence ID of our CPU in this table. Even if this table
539 * doesn't contain a patch for the CPU, scan through the whole container
540 * so that it can be skipped in case there are other containers appended.
541 */
542 eq_id = find_equiv_id(&table, bsp_cpuid_1_eax);
543
544 buf += hdr[2] + CONTAINER_HDR_SZ;
545 size -= hdr[2] + CONTAINER_HDR_SZ;
546
547 /*
548 * Scan through the rest of the container to find where it ends. We do
549 * some basic sanity-checking too.
550 */
551 while (size > 0) {
552 struct microcode_amd *mc;
553 u32 patch_size;
554 int ret;
555
556 ret = verify_patch(buf, size, &patch_size);
557 if (ret < 0) {
558 /*
559 * Patch verification failed, skip to the next container, if
560 * there is one. Before exit, check whether that container has
561 * found a patch already. If so, use it.
562 */
563 goto out;
564 } else if (ret > 0) {
565 goto skip;
566 }
567
568 mc = (struct microcode_amd *)(buf + SECTION_HDR_SIZE);
569 if (mc_patch_matches(mc, eq_id)) {
570 desc->psize = patch_size;
571 desc->mc = mc;
572 }
573
574 skip:
575 /* Skip patch section header too: */
576 buf += patch_size + SECTION_HDR_SIZE;
577 size -= patch_size + SECTION_HDR_SIZE;
578 }
579
580 out:
581 /*
582 * If we have found a patch (desc->mc), it means we're looking at the
583 * container which has a patch for this CPU so return 0 to mean, @ucode
584 * already points to the proper container. Otherwise, we return the size
585 * we scanned so that we can advance to the next container in the
586 * buffer.
587 */
588 if (desc->mc) {
589 desc->data = ucode;
590 desc->size = orig_size - size;
591
592 return 0;
593 }
594
595 return orig_size - size;
596 }
597
598 /*
599 * Scan the ucode blob for the proper container as we can have multiple
600 * containers glued together.
601 */
scan_containers(u8 * ucode,size_t size,struct cont_desc * desc)602 static void scan_containers(u8 *ucode, size_t size, struct cont_desc *desc)
603 {
604 while (size) {
605 size_t s = parse_container(ucode, size, desc);
606 if (!s)
607 return;
608
609 /* catch wraparound */
610 if (size >= s) {
611 ucode += s;
612 size -= s;
613 } else {
614 return;
615 }
616 }
617 }
618
__apply_microcode_amd(struct microcode_amd * mc,u32 * cur_rev,unsigned int psize)619 static bool __apply_microcode_amd(struct microcode_amd *mc, u32 *cur_rev,
620 unsigned int psize)
621 {
622 unsigned long p_addr = (unsigned long)&mc->hdr.data_code;
623
624 if (!verify_sha256_digest(mc->hdr.patch_id, *cur_rev, (const u8 *)p_addr, psize))
625 return false;
626
627 native_wrmsrq(MSR_AMD64_PATCH_LOADER, p_addr);
628
629 if (x86_family(bsp_cpuid_1_eax) == 0x17) {
630 unsigned long p_addr_end = p_addr + psize - 1;
631
632 invlpg(p_addr);
633
634 /*
635 * Flush next page too if patch image is crossing a page
636 * boundary.
637 */
638 if (p_addr >> PAGE_SHIFT != p_addr_end >> PAGE_SHIFT)
639 invlpg(p_addr_end);
640 }
641
642 /* verify patch application was successful */
643 *cur_rev = get_patch_level();
644 if (*cur_rev != mc->hdr.patch_id)
645 return false;
646
647 return true;
648 }
649
get_builtin_microcode(struct cpio_data * cp)650 static bool get_builtin_microcode(struct cpio_data *cp)
651 {
652 char fw_name[36] = "amd-ucode/microcode_amd.bin";
653 u8 family = x86_family(bsp_cpuid_1_eax);
654 struct firmware fw;
655
656 if (IS_ENABLED(CONFIG_X86_32))
657 return false;
658
659 if (family >= 0x15)
660 snprintf(fw_name, sizeof(fw_name),
661 "amd-ucode/microcode_amd_fam%02hhxh.bin", family);
662
663 if (firmware_request_builtin(&fw, fw_name)) {
664 cp->size = fw.size;
665 cp->data = (void *)fw.data;
666 return true;
667 }
668
669 return false;
670 }
671
find_blobs_in_containers(struct cpio_data * ret)672 static bool __init find_blobs_in_containers(struct cpio_data *ret)
673 {
674 struct cpio_data cp;
675 bool found;
676
677 if (!get_builtin_microcode(&cp))
678 cp = find_microcode_in_initrd(ucode_path);
679
680 found = cp.data && cp.size;
681 if (found)
682 *ret = cp;
683
684 return found;
685 }
686
687 /*
688 * Early load occurs before we can vmalloc(). So we look for the microcode
689 * patch container file in initrd, traverse equivalent cpu table, look for a
690 * matching microcode patch, and update, all in initrd memory in place.
691 * When vmalloc() is available for use later -- on 64-bit during first AP load,
692 * and on 32-bit during save_microcode_in_initrd() -- we can call
693 * load_microcode_amd() to save equivalent cpu table and microcode patches in
694 * kernel heap memory.
695 */
load_ucode_amd_bsp(struct early_load_data * ed,unsigned int cpuid_1_eax)696 void __init load_ucode_amd_bsp(struct early_load_data *ed, unsigned int cpuid_1_eax)
697 {
698 struct cont_desc desc = { };
699 struct microcode_amd *mc;
700 struct cpio_data cp = { };
701 char buf[4];
702 u32 rev;
703
704 if (cmdline_find_option(boot_command_line, "microcode.amd_sha_check", buf, 4)) {
705 if (!strncmp(buf, "off", 3)) {
706 sha_check = false;
707 pr_warn_once("It is a very very bad idea to disable the blobs SHA check!\n");
708 add_taint(TAINT_CPU_OUT_OF_SPEC, LOCKDEP_STILL_OK);
709 }
710 }
711
712 bsp_cpuid_1_eax = cpuid_1_eax;
713
714 rev = get_patch_level();
715 ed->old_rev = rev;
716
717 /* Needed in load_microcode_amd() */
718 ucode_cpu_info[0].cpu_sig.sig = cpuid_1_eax;
719
720 if (!find_blobs_in_containers(&cp))
721 return;
722
723 scan_containers(cp.data, cp.size, &desc);
724
725 mc = desc.mc;
726 if (!mc)
727 return;
728
729 /*
730 * Allow application of the same revision to pick up SMT-specific
731 * changes even if the revision of the other SMT thread is already
732 * up-to-date.
733 */
734 if (ed->old_rev > mc->hdr.patch_id)
735 return;
736
737 if (__apply_microcode_amd(mc, &rev, desc.psize))
738 ed->new_rev = rev;
739 }
740
patch_cpus_equivalent(struct ucode_patch * p,struct ucode_patch * n,bool ignore_stepping)741 static inline bool patch_cpus_equivalent(struct ucode_patch *p,
742 struct ucode_patch *n,
743 bool ignore_stepping)
744 {
745 /* Zen and newer hardcode the f/m/s in the patch ID */
746 if (x86_family(bsp_cpuid_1_eax) >= 0x17) {
747 union cpuid_1_eax p_cid = ucode_rev_to_cpuid(p->patch_id);
748 union cpuid_1_eax n_cid = ucode_rev_to_cpuid(n->patch_id);
749
750 if (ignore_stepping) {
751 p_cid.stepping = 0;
752 n_cid.stepping = 0;
753 }
754
755 return p_cid.full == n_cid.full;
756 } else {
757 return p->equiv_cpu == n->equiv_cpu;
758 }
759 }
760
761 /*
762 * a small, trivial cache of per-family ucode patches
763 */
cache_find_patch(struct ucode_cpu_info * uci,u16 equiv_cpu)764 static struct ucode_patch *cache_find_patch(struct ucode_cpu_info *uci, u16 equiv_cpu)
765 {
766 struct ucode_patch *p;
767 struct ucode_patch n;
768
769 n.equiv_cpu = equiv_cpu;
770 n.patch_id = uci->cpu_sig.rev;
771
772 list_for_each_entry(p, µcode_cache, plist)
773 if (patch_cpus_equivalent(p, &n, false))
774 return p;
775
776 return NULL;
777 }
778
patch_newer(struct ucode_patch * p,struct ucode_patch * n)779 static inline int patch_newer(struct ucode_patch *p, struct ucode_patch *n)
780 {
781 /* Zen and newer hardcode the f/m/s in the patch ID */
782 if (x86_family(bsp_cpuid_1_eax) >= 0x17) {
783 union zen_patch_rev zp, zn;
784
785 zp.ucode_rev = p->patch_id;
786 zn.ucode_rev = n->patch_id;
787
788 if (zn.stepping != zp.stepping)
789 return -1;
790
791 return zn.rev > zp.rev;
792 } else {
793 return n->patch_id > p->patch_id;
794 }
795 }
796
update_cache(struct ucode_patch * new_patch)797 static void update_cache(struct ucode_patch *new_patch)
798 {
799 struct ucode_patch *p;
800 int ret;
801
802 list_for_each_entry(p, µcode_cache, plist) {
803 if (patch_cpus_equivalent(p, new_patch, true)) {
804 ret = patch_newer(p, new_patch);
805 if (ret < 0)
806 continue;
807 else if (!ret) {
808 /* we already have the latest patch */
809 kfree(new_patch->data);
810 kfree(new_patch);
811 return;
812 }
813
814 list_replace(&p->plist, &new_patch->plist);
815 kfree(p->data);
816 kfree(p);
817 return;
818 }
819 }
820 /* no patch found, add it */
821 list_add_tail(&new_patch->plist, µcode_cache);
822 }
823
free_cache(void)824 static void free_cache(void)
825 {
826 struct ucode_patch *p, *tmp;
827
828 list_for_each_entry_safe(p, tmp, µcode_cache, plist) {
829 __list_del(p->plist.prev, p->plist.next);
830 kfree(p->data);
831 kfree(p);
832 }
833 }
834
find_patch(unsigned int cpu)835 static struct ucode_patch *find_patch(unsigned int cpu)
836 {
837 struct ucode_cpu_info *uci = ucode_cpu_info + cpu;
838 u16 equiv_id = 0;
839
840 uci->cpu_sig.rev = get_patch_level();
841
842 if (x86_family(bsp_cpuid_1_eax) < 0x17) {
843 equiv_id = find_equiv_id(&equiv_table, uci->cpu_sig.sig);
844 if (!equiv_id)
845 return NULL;
846 }
847
848 return cache_find_patch(uci, equiv_id);
849 }
850
reload_ucode_amd(unsigned int cpu)851 void reload_ucode_amd(unsigned int cpu)
852 {
853 u32 rev, dummy __always_unused;
854 struct microcode_amd *mc;
855 struct ucode_patch *p;
856
857 p = find_patch(cpu);
858 if (!p)
859 return;
860
861 mc = p->data;
862
863 rev = get_patch_level();
864 if (rev < mc->hdr.patch_id) {
865 if (__apply_microcode_amd(mc, &rev, p->size))
866 pr_info_once("reload revision: 0x%08x\n", rev);
867 }
868 }
869
collect_cpu_info_amd(int cpu,struct cpu_signature * csig)870 static int collect_cpu_info_amd(int cpu, struct cpu_signature *csig)
871 {
872 struct ucode_cpu_info *uci = ucode_cpu_info + cpu;
873 struct ucode_patch *p;
874
875 csig->sig = cpuid_eax(0x00000001);
876 csig->rev = get_patch_level();
877
878 /*
879 * a patch could have been loaded early, set uci->mc so that
880 * mc_bp_resume() can call apply_microcode()
881 */
882 p = find_patch(cpu);
883 if (p && (p->patch_id == csig->rev))
884 uci->mc = p->data;
885
886 return 0;
887 }
888
apply_microcode_amd(int cpu)889 static enum ucode_state apply_microcode_amd(int cpu)
890 {
891 struct cpuinfo_x86 *c = &cpu_data(cpu);
892 struct microcode_amd *mc_amd;
893 struct ucode_cpu_info *uci;
894 struct ucode_patch *p;
895 enum ucode_state ret;
896 u32 rev;
897
898 BUG_ON(raw_smp_processor_id() != cpu);
899
900 uci = ucode_cpu_info + cpu;
901
902 p = find_patch(cpu);
903 if (!p)
904 return UCODE_NFOUND;
905
906 rev = uci->cpu_sig.rev;
907
908 mc_amd = p->data;
909 uci->mc = p->data;
910
911 /* need to apply patch? */
912 if (rev > mc_amd->hdr.patch_id) {
913 ret = UCODE_OK;
914 goto out;
915 }
916
917 if (!__apply_microcode_amd(mc_amd, &rev, p->size)) {
918 pr_err("CPU%d: update failed for patch_level=0x%08x\n",
919 cpu, mc_amd->hdr.patch_id);
920 return UCODE_ERROR;
921 }
922
923 rev = mc_amd->hdr.patch_id;
924 ret = UCODE_UPDATED;
925
926 out:
927 uci->cpu_sig.rev = rev;
928 c->microcode = rev;
929
930 /* Update boot_cpu_data's revision too, if we're on the BSP: */
931 if (c->cpu_index == boot_cpu_data.cpu_index)
932 boot_cpu_data.microcode = rev;
933
934 return ret;
935 }
936
load_ucode_amd_ap(unsigned int cpuid_1_eax)937 void load_ucode_amd_ap(unsigned int cpuid_1_eax)
938 {
939 unsigned int cpu = smp_processor_id();
940
941 ucode_cpu_info[cpu].cpu_sig.sig = cpuid_1_eax;
942 apply_microcode_amd(cpu);
943 }
944
install_equiv_cpu_table(const u8 * buf,size_t buf_size)945 static size_t install_equiv_cpu_table(const u8 *buf, size_t buf_size)
946 {
947 u32 equiv_tbl_len;
948 const u32 *hdr;
949
950 if (!verify_equivalence_table(buf, buf_size))
951 return 0;
952
953 hdr = (const u32 *)buf;
954 equiv_tbl_len = hdr[2];
955
956 /* Zen and newer do not need an equivalence table. */
957 if (x86_family(bsp_cpuid_1_eax) >= 0x17)
958 goto out;
959
960 equiv_table.entry = vmalloc(equiv_tbl_len);
961 if (!equiv_table.entry) {
962 pr_err("failed to allocate equivalent CPU table\n");
963 return 0;
964 }
965
966 memcpy(equiv_table.entry, buf + CONTAINER_HDR_SZ, equiv_tbl_len);
967 equiv_table.num_entries = equiv_tbl_len / sizeof(struct equiv_cpu_entry);
968
969 out:
970 /* add header length */
971 return equiv_tbl_len + CONTAINER_HDR_SZ;
972 }
973
free_equiv_cpu_table(void)974 static void free_equiv_cpu_table(void)
975 {
976 if (x86_family(bsp_cpuid_1_eax) >= 0x17)
977 return;
978
979 vfree(equiv_table.entry);
980 memset(&equiv_table, 0, sizeof(equiv_table));
981 }
982
cleanup(void)983 static void cleanup(void)
984 {
985 free_equiv_cpu_table();
986 free_cache();
987 }
988
989 /*
990 * Return a non-negative value even if some of the checks failed so that
991 * we can skip over the next patch. If we return a negative value, we
992 * signal a grave error like a memory allocation has failed and the
993 * driver cannot continue functioning normally. In such cases, we tear
994 * down everything we've used up so far and exit.
995 */
verify_and_add_patch(u8 family,u8 * fw,unsigned int leftover,unsigned int * patch_size)996 static int verify_and_add_patch(u8 family, u8 *fw, unsigned int leftover,
997 unsigned int *patch_size)
998 {
999 struct microcode_header_amd *mc_hdr;
1000 struct ucode_patch *patch;
1001 u16 proc_id;
1002 int ret;
1003
1004 ret = verify_patch(fw, leftover, patch_size);
1005 if (ret)
1006 return ret;
1007
1008 patch = kzalloc(sizeof(*patch), GFP_KERNEL);
1009 if (!patch) {
1010 pr_err("Patch allocation failure.\n");
1011 return -EINVAL;
1012 }
1013
1014 patch->data = kmemdup(fw + SECTION_HDR_SIZE, *patch_size, GFP_KERNEL);
1015 if (!patch->data) {
1016 pr_err("Patch data allocation failure.\n");
1017 kfree(patch);
1018 return -EINVAL;
1019 }
1020 patch->size = *patch_size;
1021
1022 mc_hdr = (struct microcode_header_amd *)(fw + SECTION_HDR_SIZE);
1023 proc_id = mc_hdr->processor_rev_id;
1024
1025 INIT_LIST_HEAD(&patch->plist);
1026 patch->patch_id = mc_hdr->patch_id;
1027 patch->equiv_cpu = proc_id;
1028
1029 pr_debug("%s: Adding patch_id: 0x%08x, proc_id: 0x%04x\n",
1030 __func__, patch->patch_id, proc_id);
1031
1032 /* ... and add to cache. */
1033 update_cache(patch);
1034
1035 return 0;
1036 }
1037
1038 /* Scan the blob in @data and add microcode patches to the cache. */
__load_microcode_amd(u8 family,const u8 * data,size_t size)1039 static enum ucode_state __load_microcode_amd(u8 family, const u8 *data, size_t size)
1040 {
1041 u8 *fw = (u8 *)data;
1042 size_t offset;
1043
1044 offset = install_equiv_cpu_table(data, size);
1045 if (!offset)
1046 return UCODE_ERROR;
1047
1048 fw += offset;
1049 size -= offset;
1050
1051 if (*(u32 *)fw != UCODE_UCODE_TYPE) {
1052 pr_err("invalid type field in container file section header\n");
1053 free_equiv_cpu_table();
1054 return UCODE_ERROR;
1055 }
1056
1057 while (size > 0) {
1058 unsigned int crnt_size = 0;
1059 int ret;
1060
1061 ret = verify_and_add_patch(family, fw, size, &crnt_size);
1062 if (ret < 0)
1063 return UCODE_ERROR;
1064
1065 fw += crnt_size + SECTION_HDR_SIZE;
1066 size -= (crnt_size + SECTION_HDR_SIZE);
1067 }
1068
1069 return UCODE_OK;
1070 }
1071
_load_microcode_amd(u8 family,const u8 * data,size_t size)1072 static enum ucode_state _load_microcode_amd(u8 family, const u8 *data, size_t size)
1073 {
1074 enum ucode_state ret;
1075
1076 /* free old equiv table */
1077 free_equiv_cpu_table();
1078
1079 ret = __load_microcode_amd(family, data, size);
1080 if (ret != UCODE_OK)
1081 cleanup();
1082
1083 return ret;
1084 }
1085
load_microcode_amd(u8 family,const u8 * data,size_t size)1086 static enum ucode_state load_microcode_amd(u8 family, const u8 *data, size_t size)
1087 {
1088 struct cpuinfo_x86 *c;
1089 unsigned int nid, cpu;
1090 struct ucode_patch *p;
1091 enum ucode_state ret;
1092
1093 ret = _load_microcode_amd(family, data, size);
1094 if (ret != UCODE_OK)
1095 return ret;
1096
1097 for_each_node_with_cpus(nid) {
1098 cpu = cpumask_first(cpumask_of_node(nid));
1099 c = &cpu_data(cpu);
1100
1101 p = find_patch(cpu);
1102 if (!p)
1103 continue;
1104
1105 if (c->microcode >= p->patch_id)
1106 continue;
1107
1108 ret = UCODE_NEW;
1109 }
1110
1111 return ret;
1112 }
1113
save_microcode_in_initrd(void)1114 static int __init save_microcode_in_initrd(void)
1115 {
1116 struct cpuinfo_x86 *c = &boot_cpu_data;
1117 struct cont_desc desc = { 0 };
1118 unsigned int cpuid_1_eax;
1119 enum ucode_state ret;
1120 struct cpio_data cp;
1121
1122 if (microcode_loader_disabled() || c->x86_vendor != X86_VENDOR_AMD || c->x86 < 0x10)
1123 return 0;
1124
1125 cpuid_1_eax = native_cpuid_eax(1);
1126
1127 if (!find_blobs_in_containers(&cp))
1128 return -EINVAL;
1129
1130 scan_containers(cp.data, cp.size, &desc);
1131 if (!desc.mc)
1132 return -EINVAL;
1133
1134 ret = _load_microcode_amd(x86_family(cpuid_1_eax), desc.data, desc.size);
1135 if (ret > UCODE_UPDATED)
1136 return -EINVAL;
1137
1138 return 0;
1139 }
1140 early_initcall(save_microcode_in_initrd);
1141
1142 /*
1143 * AMD microcode firmware naming convention, up to family 15h they are in
1144 * the legacy file:
1145 *
1146 * amd-ucode/microcode_amd.bin
1147 *
1148 * This legacy file is always smaller than 2K in size.
1149 *
1150 * Beginning with family 15h, they are in family-specific firmware files:
1151 *
1152 * amd-ucode/microcode_amd_fam15h.bin
1153 * amd-ucode/microcode_amd_fam16h.bin
1154 * ...
1155 *
1156 * These might be larger than 2K.
1157 */
request_microcode_amd(int cpu,struct device * device)1158 static enum ucode_state request_microcode_amd(int cpu, struct device *device)
1159 {
1160 char fw_name[36] = "amd-ucode/microcode_amd.bin";
1161 struct cpuinfo_x86 *c = &cpu_data(cpu);
1162 enum ucode_state ret = UCODE_NFOUND;
1163 const struct firmware *fw;
1164
1165 if (force_minrev)
1166 return UCODE_NFOUND;
1167
1168 if (c->x86 >= 0x15)
1169 snprintf(fw_name, sizeof(fw_name), "amd-ucode/microcode_amd_fam%.2xh.bin", c->x86);
1170
1171 if (request_firmware_direct(&fw, (const char *)fw_name, device)) {
1172 pr_debug("failed to load file %s\n", fw_name);
1173 goto out;
1174 }
1175
1176 ret = UCODE_ERROR;
1177 if (!verify_container(fw->data, fw->size))
1178 goto fw_release;
1179
1180 ret = load_microcode_amd(c->x86, fw->data, fw->size);
1181
1182 fw_release:
1183 release_firmware(fw);
1184
1185 out:
1186 return ret;
1187 }
1188
microcode_fini_cpu_amd(int cpu)1189 static void microcode_fini_cpu_amd(int cpu)
1190 {
1191 struct ucode_cpu_info *uci = ucode_cpu_info + cpu;
1192
1193 uci->mc = NULL;
1194 }
1195
finalize_late_load_amd(int result)1196 static void finalize_late_load_amd(int result)
1197 {
1198 if (result)
1199 cleanup();
1200 }
1201
1202 static struct microcode_ops microcode_amd_ops = {
1203 .request_microcode_fw = request_microcode_amd,
1204 .collect_cpu_info = collect_cpu_info_amd,
1205 .apply_microcode = apply_microcode_amd,
1206 .microcode_fini_cpu = microcode_fini_cpu_amd,
1207 .finalize_late_load = finalize_late_load_amd,
1208 .nmi_safe = true,
1209 };
1210
init_amd_microcode(void)1211 struct microcode_ops * __init init_amd_microcode(void)
1212 {
1213 struct cpuinfo_x86 *c = &boot_cpu_data;
1214
1215 if (c->x86_vendor != X86_VENDOR_AMD || c->x86 < 0x10) {
1216 pr_warn("AMD CPU family 0x%x not supported\n", c->x86);
1217 return NULL;
1218 }
1219 return µcode_amd_ops;
1220 }
1221
exit_amd_microcode(void)1222 void __exit exit_amd_microcode(void)
1223 {
1224 cleanup();
1225 }
1226