1 files changed, 346 insertions, 115 deletions

diff --git a/arch/x86/kernel/cpu/microcode/amd.c b/arch/x86/kernel/cpu/microcode/amd.c
index 31a73715d755..3821a985f4ff 100644
--- a/arch/x86/kernel/cpu/microcode/amd.c
+++ b/arch/x86/kernel/cpu/microcode/amd.c
@@ -23,17 +23,22 @@
 
 #include <linux/earlycpio.h>
 #include <linux/firmware.h>
+#include <linux/bsearch.h>
 #include <linux/uaccess.h>
 #include <linux/vmalloc.h>
 #include <linux/initrd.h>
 #include <linux/kernel.h>
 #include <linux/pci.h>
 
+#include <crypto/sha2.h>
+
 #include <asm/microcode.h>
 #include <asm/processor.h>
+#include <asm/cmdline.h>
 #include <asm/setup.h>
 #include <asm/cpu.h>
 #include <asm/msr.h>
+#include <asm/tlb.h>
 
 #include "internal.h"
 
@@ -144,6 +149,158 @@ ucode_path[] __maybe_unused = "kernel/x86/microcode/AuthenticAMD.bin";
  */
 static u32 bsp_cpuid_1_eax __ro_after_init;
 
+static bool sha_check = true;
+
+struct patch_digest {
+	u32 patch_id;
+	u8 sha256[SHA256_DIGEST_SIZE];
+};
+
+#include "amd_shas.c"
+
+static int cmp_id(const void *key, const void *elem)
+{
+	struct patch_digest *pd = (struct patch_digest *)elem;
+	u32 patch_id = *(u32 *)key;
+
+	if (patch_id == pd->patch_id)
+		return 0;
+	else if (patch_id < pd->patch_id)
+		return -1;
+	else
+		return 1;
+}
+
+static u32 cpuid_to_ucode_rev(unsigned int val)
+{
+	union zen_patch_rev p = {};
+	union cpuid_1_eax c;
+
+	c.full = val;
+
+	p.stepping  = c.stepping;
+	p.model     = c.model;
+	p.ext_model = c.ext_model;
+	p.ext_fam   = c.ext_fam;
+
+	return p.ucode_rev;
+}
+
+static u32 get_cutoff_revision(u32 rev)
+{
+	switch (rev >> 8) {
+	case 0x80012: return 0x8001277; break;
+	case 0x80082: return 0x800820f; break;
+	case 0x83010: return 0x830107c; break;
+	case 0x86001: return 0x860010e; break;
+	case 0x86081: return 0x8608108; break;
+	case 0x87010: return 0x8701034; break;
+	case 0x8a000: return 0x8a0000a; break;
+	case 0xa0010: return 0xa00107a; break;
+	case 0xa0011: return 0xa0011da; break;
+	case 0xa0012: return 0xa001243; break;
+	case 0xa0082: return 0xa00820e; break;
+	case 0xa1011: return 0xa101153; break;
+	case 0xa1012: return 0xa10124e; break;
+	case 0xa1081: return 0xa108109; break;
+	case 0xa2010: return 0xa20102f; break;
+	case 0xa2012: return 0xa201212; break;
+	case 0xa4041: return 0xa404109; break;
+	case 0xa5000: return 0xa500013; break;
+	case 0xa6012: return 0xa60120a; break;
+	case 0xa7041: return 0xa704109; break;
+	case 0xa7052: return 0xa705208; break;
+	case 0xa7080: return 0xa708009; break;
+	case 0xa70c0: return 0xa70C009; break;
+	case 0xaa001: return 0xaa00116; break;
+	case 0xaa002: return 0xaa00218; break;
+	case 0xb0021: return 0xb002146; break;
+	case 0xb0081: return 0xb008111; break;
+	case 0xb1010: return 0xb101046; break;
+	case 0xb2040: return 0xb204031; break;
+	case 0xb4040: return 0xb404031; break;
+	case 0xb4041: return 0xb404101; break;
+	case 0xb6000: return 0xb600031; break;
+	case 0xb6080: return 0xb608031; break;
+	case 0xb7000: return 0xb700031; break;
+	default: break;
+
+	}
+	return 0;
+}
+
+static bool need_sha_check(u32 cur_rev)
+{
+	u32 cutoff;
+
+	if (!cur_rev) {
+		cur_rev = cpuid_to_ucode_rev(bsp_cpuid_1_eax);
+		pr_info_once("No current revision, generating the lowest one: 0x%x\n", cur_rev);
+	}
+
+	cutoff = get_cutoff_revision(cur_rev);
+	if (cutoff)
+		return cur_rev <= cutoff;
+
+	pr_info("You should not be seeing this. Please send the following couple of lines to x86-<at>-kernel.org\n");
+	pr_info("CPUID(1).EAX: 0x%x, current revision: 0x%x\n", bsp_cpuid_1_eax, cur_rev);
+	return true;
+}
+
+static bool cpu_has_entrysign(void)
+{
+	unsigned int fam   = x86_family(bsp_cpuid_1_eax);
+	unsigned int model = x86_model(bsp_cpuid_1_eax);
+
+	if (fam == 0x17 || fam == 0x19)
+		return true;
+
+	if (fam == 0x1a) {
+		if (model <= 0x2f ||
+		    (0x40 <= model && model <= 0x4f) ||
+		    (0x60 <= model && model <= 0x6f))
+			return true;
+	}
+
+	return false;
+}
+
+static bool verify_sha256_digest(u32 patch_id, u32 cur_rev, const u8 *data, unsigned int len)
+{
+	struct patch_digest *pd = NULL;
+	u8 digest[SHA256_DIGEST_SIZE];
+	int i;
+
+	if (!cpu_has_entrysign())
+		return true;
+
+	if (!need_sha_check(cur_rev))
+		return true;
+
+	if (!sha_check)
+		return true;
+
+	pd = bsearch(&patch_id, phashes, ARRAY_SIZE(phashes), sizeof(struct patch_digest), cmp_id);
+	if (!pd) {
+		pr_err("No sha256 digest for patch ID: 0x%x found\n", patch_id);
+		return false;
+	}
+
+	sha256(data, len, digest);
+
+	if (memcmp(digest, pd->sha256, sizeof(digest))) {
+		pr_err("Patch 0x%x SHA256 digest mismatch!\n", patch_id);
+
+		for (i = 0; i < SHA256_DIGEST_SIZE; i++)
+			pr_cont("0x%x ", digest[i]);
+		pr_info("\n");
+
+		return false;
+	}
+
+	return true;
+}
+
 static union cpuid_1_eax ucode_rev_to_cpuid(unsigned int val)
 {
 	union zen_patch_rev p;
@@ -161,6 +318,30 @@ static union cpuid_1_eax ucode_rev_to_cpuid(unsigned int val)
 	return c;
 }
 
+static u32 get_patch_level(void)
+{
+	u32 rev, dummy __always_unused;
+
+	if (IS_ENABLED(CONFIG_MICROCODE_DBG)) {
+		int cpu = smp_processor_id();
+
+		if (!microcode_rev[cpu]) {
+			if (!base_rev)
+				base_rev = cpuid_to_ucode_rev(bsp_cpuid_1_eax);
+
+			microcode_rev[cpu] = base_rev;
+
+			ucode_dbg("CPU%d, base_rev: 0x%x\n", cpu, base_rev);
+		}
+
+		return microcode_rev[cpu];
+	}
+
+	native_rdmsr(MSR_AMD64_PATCH_LEVEL, rev, dummy);
+
+	return rev;
+}
+
 static u16 find_equiv_id(struct equiv_cpu_table *et, u32 sig)
 {
 	unsigned int i;
@@ -190,13 +371,13 @@ static bool verify_container(const u8 *buf, size_t buf_size)
 	u32 cont_magic;
 
 	if (buf_size <= CONTAINER_HDR_SZ) {
-		pr_debug("Truncated microcode container header.\n");
+		ucode_dbg("Truncated microcode container header.\n");
 		return false;
 	}
 
 	cont_magic = *(const u32 *)buf;
 	if (cont_magic != UCODE_MAGIC) {
-		pr_debug("Invalid magic value (0x%08x).\n", cont_magic);
+		ucode_dbg("Invalid magic value (0x%08x).\n", cont_magic);
 		return false;
 	}
 
@@ -221,8 +402,8 @@ static bool verify_equivalence_table(const u8 *buf, size_t buf_size)
 
 	cont_type = hdr[1];
 	if (cont_type != UCODE_EQUIV_CPU_TABLE_TYPE) {
-		pr_debug("Wrong microcode container equivalence table type: %u.\n",
-			 cont_type);
+		ucode_dbg("Wrong microcode container equivalence table type: %u.\n",
+			  cont_type);
 		return false;
 	}
 
@@ -231,7 +412,7 @@ static bool verify_equivalence_table(const u8 *buf, size_t buf_size)
 	equiv_tbl_len = hdr[2];
 	if (equiv_tbl_len < sizeof(struct equiv_cpu_entry) ||
 	    buf_size < equiv_tbl_len) {
-		pr_debug("Truncated equivalence table.\n");
+		ucode_dbg("Truncated equivalence table.\n");
 		return false;
 	}
 
@@ -245,14 +426,13 @@ static bool verify_equivalence_table(const u8 *buf, size_t buf_size)
  * On success, @sh_psize returns the patch size according to the section header,
  * to the caller.
  */
-static bool
-__verify_patch_section(const u8 *buf, size_t buf_size, u32 *sh_psize)
+static bool __verify_patch_section(const u8 *buf, size_t buf_size, u32 *sh_psize)
 {
 	u32 p_type, p_size;
 	const u32 *hdr;
 
 	if (buf_size < SECTION_HDR_SIZE) {
-		pr_debug("Truncated patch section.\n");
+		ucode_dbg("Truncated patch section.\n");
 		return false;
 	}
 
@@ -261,13 +441,13 @@ __verify_patch_section(const u8 *buf, size_t buf_size, u32 *sh_psize)
 	p_size = hdr[1];
 
 	if (p_type != UCODE_UCODE_TYPE) {
-		pr_debug("Invalid type field (0x%x) in container file section header.\n",
-			 p_type);
+		ucode_dbg("Invalid type field (0x%x) in container file section header.\n",
+			  p_type);
 		return false;
 	}
 
 	if (p_size < sizeof(struct microcode_header_amd)) {
-		pr_debug("Patch of size %u too short.\n", p_size);
+		ucode_dbg("Patch of size %u too short.\n", p_size);
 		return false;
 	}
 
@@ -282,13 +462,13 @@ __verify_patch_section(const u8 *buf, size_t buf_size, u32 *sh_psize)
  * exceed the per-family maximum). @sh_psize is the size read from the section
  * header.
  */
-static unsigned int __verify_patch_size(u32 sh_psize, size_t buf_size)
+static bool __verify_patch_size(u32 sh_psize, size_t buf_size)
 {
 	u8 family = x86_family(bsp_cpuid_1_eax);
 	u32 max_size;
 
 	if (family >= 0x15)
-		return min_t(u32, sh_psize, buf_size);
+		goto ret;
 
 #define F1XH_MPB_MAX_SIZE 2048
 #define F14H_MPB_MAX_SIZE 1824
@@ -302,13 +482,15 @@ static unsigned int __verify_patch_size(u32 sh_psize, size_t buf_size)
 		break;
 	default:
 		WARN(1, "%s: WTF family: 0x%x\n", __func__, family);
-		return 0;
+		return false;
 	}
 
-	if (sh_psize > min_t(u32, buf_size, max_size))
-		return 0;
+	if (sh_psize > max_size)
+		return false;
 
-	return sh_psize;
+ret:
+	/* Working with the whole buffer so < is ok. */
+	return sh_psize <= buf_size;
 }
 
 /*
@@ -323,7 +505,7 @@ static int verify_patch(const u8 *buf, size_t buf_size, u32 *patch_size)
 {
 	u8 family = x86_family(bsp_cpuid_1_eax);
 	struct microcode_header_amd *mc_hdr;
-	unsigned int ret;
+	u32 cur_rev, cutoff, patch_rev;
 	u32 sh_psize;
 	u16 proc_id;
 	u8 patch_fam;
@@ -343,13 +525,12 @@ static int verify_patch(const u8 *buf, size_t buf_size, u32 *patch_size)
 	 * size sh_psize, as the section claims.
 	 */
 	if (buf_size < sh_psize) {
-		pr_debug("Patch of size %u truncated.\n", sh_psize);
+		ucode_dbg("Patch of size %u truncated.\n", sh_psize);
 		return -1;
 	}
 
-	ret = __verify_patch_size(sh_psize, buf_size);
-	if (!ret) {
-		pr_debug("Per-family patch size mismatch.\n");
+	if (!__verify_patch_size(sh_psize, buf_size)) {
+		ucode_dbg("Per-family patch size mismatch.\n");
 		return -1;
 	}
 
@@ -363,9 +544,33 @@ static int verify_patch(const u8 *buf, size_t buf_size, u32 *patch_size)
 
 	proc_id	= mc_hdr->processor_rev_id;
 	patch_fam = 0xf + (proc_id >> 12);
+
 	if (patch_fam != family)
 		return 1;
 
+	cur_rev = get_patch_level();
+
+	/* No cutoff revision means old/unaffected by signing algorithm weakness => matches */
+	cutoff = get_cutoff_revision(cur_rev);
+	if (!cutoff)
+		goto ok;
+
+	patch_rev = mc_hdr->patch_id;
+
+	ucode_dbg("cur_rev: 0x%x, cutoff: 0x%x, patch_rev: 0x%x\n",
+		  cur_rev, cutoff, patch_rev);
+
+	if (cur_rev <= cutoff && patch_rev <= cutoff)
+		goto ok;
+
+	if (cur_rev > cutoff && patch_rev > cutoff)
+		goto ok;
+
+	return 1;
+
+ok:
+	ucode_dbg("Patch-ID 0x%08x: family: 0x%x\n", mc_hdr->patch_id, patch_fam);
+
 	return 0;
 }
 
@@ -380,8 +585,8 @@ static bool mc_patch_matches(struct microcode_amd *mc, u16 eq_id)
 
 /*
  * This scans the ucode blob for the proper container as we can have multiple
- * containers glued together. Returns the equivalence ID from the equivalence
- * table or 0 if none found.
+ * containers glued together.
+ *
  * Returns the amount of bytes consumed while scanning. @desc contains all the
  * data we're going to use in later stages of the application.
  */
@@ -433,9 +638,12 @@ static size_t parse_container(u8 *ucode, size_t size, struct cont_desc *desc)
 		}
 
 		mc = (struct microcode_amd *)(buf + SECTION_HDR_SIZE);
+
 		if (mc_patch_matches(mc, eq_id)) {
 			desc->psize = patch_size;
 			desc->mc = mc;
+
+			ucode_dbg(" match: size: %d\n", patch_size);
 		}
 
 skip:
@@ -483,53 +691,41 @@ static void scan_containers(u8 *ucode, size_t size, struct cont_desc *desc)
 	}
 }
 
-static int __apply_microcode_amd(struct microcode_amd *mc)
+static bool __apply_microcode_amd(struct microcode_amd *mc, u32 *cur_rev,
+				  unsigned int psize)
 {
-	u32 rev, dummy;
+	unsigned long p_addr = (unsigned long)&mc->hdr.data_code;
 
-	native_wrmsrl(MSR_AMD64_PATCH_LOADER, (u64)(long)&mc->hdr.data_code);
+	if (!verify_sha256_digest(mc->hdr.patch_id, *cur_rev, (const u8 *)p_addr, psize))
+		return false;
 
-	/* verify patch application was successful */
-	native_rdmsr(MSR_AMD64_PATCH_LEVEL, rev, dummy);
+	native_wrmsrq(MSR_AMD64_PATCH_LOADER, p_addr);
 
-	if (rev != mc->hdr.patch_id)
-		return -1;
+	if (x86_family(bsp_cpuid_1_eax) == 0x17) {
+		unsigned long p_addr_end = p_addr + psize - 1;
 
-	return 0;
-}
+		invlpg(p_addr);
 
-/*
- * Early load occurs before we can vmalloc(). So we look for the microcode
- * patch container file in initrd, traverse equivalent cpu table, look for a
- * matching microcode patch, and update, all in initrd memory in place.
- * When vmalloc() is available for use later -- on 64-bit during first AP load,
- * and on 32-bit during save_microcode_in_initrd_amd() -- we can call
- * load_microcode_amd() to save equivalent cpu table and microcode patches in
- * kernel heap memory.
- *
- * Returns true if container found (sets @desc), false otherwise.
- */
-static bool early_apply_microcode(u32 old_rev, void *ucode, size_t size)
-{
-	struct cont_desc desc = { 0 };
-	struct microcode_amd *mc;
-	bool ret = false;
+		/*
+		 * Flush next page too if patch image is crossing a page
+		 * boundary.
+		 */
+		if (p_addr >> PAGE_SHIFT != p_addr_end >> PAGE_SHIFT)
+			invlpg(p_addr_end);
+	}
 
-	scan_containers(ucode, size, &desc);
+	if (IS_ENABLED(CONFIG_MICROCODE_DBG))
+		microcode_rev[smp_processor_id()] = mc->hdr.patch_id;
 
-	mc = desc.mc;
-	if (!mc)
-		return ret;
+	/* verify patch application was successful */
+	*cur_rev = get_patch_level();
 
-	/*
-	 * Allow application of the same revision to pick up SMT-specific
-	 * changes even if the revision of the other SMT thread is already
-	 * up-to-date.
-	 */
-	if (old_rev > mc->hdr.patch_id)
-		return ret;
+	ucode_dbg("updated rev: 0x%x\n", *cur_rev);
+
+	if (*cur_rev != mc->hdr.patch_id)
+		return false;
 
-	return !__apply_microcode_amd(mc);
+	return true;
 }
 
 static bool get_builtin_microcode(struct cpio_data *cp)
@@ -554,64 +750,74 @@ static bool get_builtin_microcode(struct cpio_data *cp)
 	return false;
 }
 
-static void __init find_blobs_in_containers(struct cpio_data *ret)
+static bool __init find_blobs_in_containers(struct cpio_data *ret)
 {
 	struct cpio_data cp;
+	bool found;
 
 	if (!get_builtin_microcode(&cp))
 		cp = find_microcode_in_initrd(ucode_path);
 
-	*ret = cp;
+	found = cp.data && cp.size;
+	if (found)
+		*ret = cp;
+
+	return found;
 }
 
+/*
+ * Early load occurs before we can vmalloc(). So we look for the microcode
+ * patch container file in initrd, traverse equivalent cpu table, look for a
+ * matching microcode patch, and update, all in initrd memory in place.
+ * When vmalloc() is available for use later -- on 64-bit during first AP load,
+ * and on 32-bit during save_microcode_in_initrd() -- we can call
+ * load_microcode_amd() to save equivalent cpu table and microcode patches in
+ * kernel heap memory.
+ */
 void __init load_ucode_amd_bsp(struct early_load_data *ed, unsigned int cpuid_1_eax)
 {
+	struct cont_desc desc = { };
+	struct microcode_amd *mc;
 	struct cpio_data cp = { };
-	u32 dummy;
+	char buf[4];
+	u32 rev;
+
+	if (cmdline_find_option(boot_command_line, "microcode.amd_sha_check", buf, 4)) {
+		if (!strncmp(buf, "off", 3)) {
+			sha_check = false;
+			pr_warn_once("It is a very very bad idea to disable the blobs SHA check!\n");
+			add_taint(TAINT_CPU_OUT_OF_SPEC, LOCKDEP_STILL_OK);
+		}
+	}
 
 	bsp_cpuid_1_eax = cpuid_1_eax;
 
-	native_rdmsr(MSR_AMD64_PATCH_LEVEL, ed->old_rev, dummy);
+	rev = get_patch_level();
+	ed->old_rev = rev;
 
 	/* Needed in load_microcode_amd() */
 	ucode_cpu_info[0].cpu_sig.sig = cpuid_1_eax;
 
-	find_blobs_in_containers(&cp);
-	if (!(cp.data && cp.size))
+	if (!find_blobs_in_containers(&cp))
 		return;
 
-	if (early_apply_microcode(ed->old_rev, cp.data, cp.size))
-		native_rdmsr(MSR_AMD64_PATCH_LEVEL, ed->new_rev, dummy);
-}
-
-static enum ucode_state _load_microcode_amd(u8 family, const u8 *data, size_t size);
-
-static int __init save_microcode_in_initrd(void)
-{
-	unsigned int cpuid_1_eax = native_cpuid_eax(1);
-	struct cpuinfo_x86 *c = &boot_cpu_data;
-	struct cont_desc desc = { 0 };
-	enum ucode_state ret;
-	struct cpio_data cp;
-
-	if (dis_ucode_ldr || c->x86_vendor != X86_VENDOR_AMD || c->x86 < 0x10)
-		return 0;
-
-	find_blobs_in_containers(&cp);
-	if (!(cp.data && cp.size))
-		return -EINVAL;
-
 	scan_containers(cp.data, cp.size, &desc);
-	if (!desc.mc)
-		return -EINVAL;
 
-	ret = _load_microcode_amd(x86_family(cpuid_1_eax), desc.data, desc.size);
-	if (ret > UCODE_UPDATED)
-		return -EINVAL;
+	mc = desc.mc;
+	if (!mc)
+		return;
 
-	return 0;
+	/*
+	 * Allow application of the same revision to pick up SMT-specific
+	 * changes even if the revision of the other SMT thread is already
+	 * up-to-date.
+	 */
+	if (ed->old_rev > mc->hdr.patch_id)
+		return;
+
+	if (__apply_microcode_amd(mc, &rev, desc.psize))
+		ed->new_rev = rev;
 }
-early_initcall(save_microcode_in_initrd);
 
 static inline bool patch_cpus_equivalent(struct ucode_patch *p,
 					 struct ucode_patch *n,
@@ -644,8 +850,6 @@ static struct ucode_patch *cache_find_patch(struct ucode_cpu_info *uci, u16 equi
 	n.equiv_cpu = equiv_cpu;
 	n.patch_id  = uci->cpu_sig.rev;
 
-	WARN_ON_ONCE(!n.patch_id);
-
 	list_for_each_entry(p, &microcode_cache, plist)
 		if (patch_cpus_equivalent(p, &n, false))
 			return p;
@@ -712,14 +916,9 @@ static void free_cache(void)
 static struct ucode_patch *find_patch(unsigned int cpu)
 {
 	struct ucode_cpu_info *uci = ucode_cpu_info + cpu;
-	u32 rev, dummy __always_unused;
 	u16 equiv_id = 0;
 
-	/* fetch rev if not populated yet: */
-	if (!uci->cpu_sig.rev) {
-		rdmsr(MSR_AMD64_PATCH_LEVEL, rev, dummy);
-		uci->cpu_sig.rev = rev;
-	}
+	uci->cpu_sig.rev = get_patch_level();
 
 	if (x86_family(bsp_cpuid_1_eax) < 0x17) {
 		equiv_id = find_equiv_id(&equiv_table, uci->cpu_sig.sig);
@@ -742,22 +941,20 @@ void reload_ucode_amd(unsigned int cpu)
 
 	mc = p->data;
 
-	rdmsr(MSR_AMD64_PATCH_LEVEL, rev, dummy);
-
+	rev = get_patch_level();
 	if (rev < mc->hdr.patch_id) {
-		if (!__apply_microcode_amd(mc))
-			pr_info_once("reload revision: 0x%08x\n", mc->hdr.patch_id);
+		if (__apply_microcode_amd(mc, &rev, p->size))
+			pr_info_once("reload revision: 0x%08x\n", rev);
 	}
 }
 
 static int collect_cpu_info_amd(int cpu, struct cpu_signature *csig)
 {
-	struct cpuinfo_x86 *c = &cpu_data(cpu);
 	struct ucode_cpu_info *uci = ucode_cpu_info + cpu;
 	struct ucode_patch *p;
 
 	csig->sig = cpuid_eax(0x00000001);
-	csig->rev = c->microcode;
+	csig->rev = get_patch_level();
 
 	/*
 	 * a patch could have been loaded early, set uci->mc so that
@@ -798,7 +995,7 @@ static enum ucode_state apply_microcode_amd(int cpu)
 		goto out;
 	}
 
-	if (__apply_microcode_amd(mc_amd)) {
+	if (!__apply_microcode_amd(mc_amd, &rev, p->size)) {
 		pr_err("CPU%d: update failed for patch_level=0x%08x\n",
 			cpu, mc_amd->hdr.patch_id);
 		return UCODE_ERROR;
@@ -910,7 +1107,7 @@ static int verify_and_add_patch(u8 family, u8 *fw, unsigned int leftover,
 	patch->patch_id  = mc_hdr->patch_id;
 	patch->equiv_cpu = proc_id;
 
-	pr_debug("%s: Adding patch_id: 0x%08x, proc_id: 0x%04x\n",
+	ucode_dbg("%s: Adding patch_id: 0x%08x, proc_id: 0x%04x\n",
 		 __func__, patch->patch_id, proc_id);
 
 	/* ... and add to cache. */
@@ -920,8 +1117,7 @@ static int verify_and_add_patch(u8 family, u8 *fw, unsigned int leftover,
 }
 
 /* Scan the blob in @data and add microcode patches to the cache. */
-static enum ucode_state __load_microcode_amd(u8 family, const u8 *data,
-					     size_t size)
+static enum ucode_state __load_microcode_amd(u8 family, const u8 *data, size_t size)
 {
 	u8 *fw = (u8 *)data;
 	size_t offset;
@@ -979,7 +1175,7 @@ static enum ucode_state load_microcode_amd(u8 family, const u8 *data, size_t siz
 	if (ret != UCODE_OK)
 		return ret;
 
-	for_each_node(nid) {
+	for_each_node_with_cpus(nid) {
 		cpu = cpumask_first(cpumask_of_node(nid));
 		c = &cpu_data(cpu);
 
@@ -996,6 +1192,34 @@ static enum ucode_state load_microcode_amd(u8 family, const u8 *data, size_t siz
 	return ret;
 }
 
+static int __init save_microcode_in_initrd(void)
+{
+	struct cpuinfo_x86 *c = &boot_cpu_data;
+	struct cont_desc desc = { 0 };
+	unsigned int cpuid_1_eax;
+	enum ucode_state ret;
+	struct cpio_data cp;
+
+	if (microcode_loader_disabled() || c->x86_vendor != X86_VENDOR_AMD || c->x86 < 0x10)
+		return 0;
+
+	cpuid_1_eax = native_cpuid_eax(1);
+
+	if (!find_blobs_in_containers(&cp))
+		return -EINVAL;
+
+	scan_containers(cp.data, cp.size, &desc);
+	if (!desc.mc)
+		return -EINVAL;
+
+	ret = _load_microcode_amd(x86_family(cpuid_1_eax), desc.data, desc.size);
+	if (ret > UCODE_UPDATED)
+		return -EINVAL;
+
+	return 0;
+}
+early_initcall(save_microcode_in_initrd);
+
 /*
  * AMD microcode firmware naming convention, up to family 15h they are in
  * the legacy file:
@@ -1026,7 +1250,7 @@ static enum ucode_state request_microcode_amd(int cpu, struct device *device)
 		snprintf(fw_name, sizeof(fw_name), "amd-ucode/microcode_amd_fam%.2xh.bin", c->x86);
 
 	if (request_firmware_direct(&fw, (const char *)fw_name, device)) {
-		pr_debug("failed to load file %s\n", fw_name);
+		ucode_dbg("failed to load file %s\n", fw_name);
 		goto out;
 	}
 
@@ -1050,11 +1274,18 @@ static void microcode_fini_cpu_amd(int cpu)
 	uci->mc = NULL;
 }
 
+static void finalize_late_load_amd(int result)
+{
+	if (result)
+		cleanup();
+}
+
 static struct microcode_ops microcode_amd_ops = {
 	.request_microcode_fw	= request_microcode_amd,
 	.collect_cpu_info	= collect_cpu_info_amd,
 	.apply_microcode	= apply_microcode_amd,
 	.microcode_fini_cpu	= microcode_fini_cpu_amd,
+	.finalize_late_load	= finalize_late_load_amd,
 	.nmi_safe		= true,
 };