xref: /linux/arch/x86/kernel/espfix_64.c (revision 8ca4fc323d2e4ab9dabbdd57633af40b0c7e6af9)
1 // SPDX-License-Identifier: GPL-2.0-only
2 /* ----------------------------------------------------------------------- *
3  *
4  *   Copyright 2014 Intel Corporation; author: H. Peter Anvin
5  *
6  * ----------------------------------------------------------------------- */
7 
8 /*
9  * The IRET instruction, when returning to a 16-bit segment, only
10  * restores the bottom 16 bits of the user space stack pointer.  This
11  * causes some 16-bit software to break, but it also leaks kernel state
12  * to user space.
13  *
14  * This works around this by creating percpu "ministacks", each of which
15  * is mapped 2^16 times 64K apart.  When we detect that the return SS is
16  * on the LDT, we copy the IRET frame to the ministack and use the
17  * relevant alias to return to userspace.  The ministacks are mapped
18  * readonly, so if the IRET fault we promote #GP to #DF which is an IST
19  * vector and thus has its own stack; we then do the fixup in the #DF
20  * handler.
21  *
22  * This file sets up the ministacks and the related page tables.  The
23  * actual ministack invocation is in entry_64.S.
24  */
25 
26 #include <linux/init.h>
27 #include <linux/init_task.h>
28 #include <linux/kernel.h>
29 #include <linux/percpu.h>
30 #include <linux/gfp.h>
31 #include <linux/random.h>
32 #include <linux/pgtable.h>
33 #include <asm/pgalloc.h>
34 #include <asm/setup.h>
35 #include <asm/espfix.h>
36 
37 /*
38  * Note: we only need 6*8 = 48 bytes for the espfix stack, but round
39  * it up to a cache line to avoid unnecessary sharing.
40  */
41 #define ESPFIX_STACK_SIZE	(8*8UL)
42 #define ESPFIX_STACKS_PER_PAGE	(PAGE_SIZE/ESPFIX_STACK_SIZE)
43 
44 /* There is address space for how many espfix pages? */
45 #define ESPFIX_PAGE_SPACE	(1UL << (P4D_SHIFT-PAGE_SHIFT-16))
46 
47 #define ESPFIX_MAX_CPUS		(ESPFIX_STACKS_PER_PAGE * ESPFIX_PAGE_SPACE)
48 #if CONFIG_NR_CPUS > ESPFIX_MAX_CPUS
49 # error "Need more virtual address space for the ESPFIX hack"
50 #endif
51 
52 #define PGALLOC_GFP (GFP_KERNEL | __GFP_ZERO)
53 
54 /* This contains the *bottom* address of the espfix stack */
55 DEFINE_PER_CPU_READ_MOSTLY(unsigned long, espfix_stack);
56 DEFINE_PER_CPU_READ_MOSTLY(unsigned long, espfix_waddr);
57 
58 /* Initialization mutex - should this be a spinlock? */
59 static DEFINE_MUTEX(espfix_init_mutex);
60 
61 /* Page allocation bitmap - each page serves ESPFIX_STACKS_PER_PAGE CPUs */
62 #define ESPFIX_MAX_PAGES  DIV_ROUND_UP(CONFIG_NR_CPUS, ESPFIX_STACKS_PER_PAGE)
63 static void *espfix_pages[ESPFIX_MAX_PAGES];
64 
65 static __page_aligned_bss pud_t espfix_pud_page[PTRS_PER_PUD]
66 	__aligned(PAGE_SIZE);
67 
68 static unsigned int page_random, slot_random;
69 
70 /*
71  * This returns the bottom address of the espfix stack for a specific CPU.
72  * The math allows for a non-power-of-two ESPFIX_STACK_SIZE, in which case
73  * we have to account for some amount of padding at the end of each page.
74  */
75 static inline unsigned long espfix_base_addr(unsigned int cpu)
76 {
77 	unsigned long page, slot;
78 	unsigned long addr;
79 
80 	page = (cpu / ESPFIX_STACKS_PER_PAGE) ^ page_random;
81 	slot = (cpu + slot_random) % ESPFIX_STACKS_PER_PAGE;
82 	addr = (page << PAGE_SHIFT) + (slot * ESPFIX_STACK_SIZE);
83 	addr = (addr & 0xffffUL) | ((addr & ~0xffffUL) << 16);
84 	addr += ESPFIX_BASE_ADDR;
85 	return addr;
86 }
87 
88 #define PTE_STRIDE        (65536/PAGE_SIZE)
89 #define ESPFIX_PTE_CLONES (PTRS_PER_PTE/PTE_STRIDE)
90 #define ESPFIX_PMD_CLONES PTRS_PER_PMD
91 #define ESPFIX_PUD_CLONES (65536/(ESPFIX_PTE_CLONES*ESPFIX_PMD_CLONES))
92 
93 #define PGTABLE_PROT	  ((_KERNPG_TABLE & ~_PAGE_RW) | _PAGE_NX)
94 
95 static void init_espfix_random(void)
96 {
97 	unsigned long rand = get_random_long();
98 
99 	slot_random = rand % ESPFIX_STACKS_PER_PAGE;
100 	page_random = (rand / ESPFIX_STACKS_PER_PAGE)
101 		& (ESPFIX_PAGE_SPACE - 1);
102 }
103 
104 void __init init_espfix_bsp(void)
105 {
106 	pgd_t *pgd;
107 	p4d_t *p4d;
108 
109 	/* Install the espfix pud into the kernel page directory */
110 	pgd = &init_top_pgt[pgd_index(ESPFIX_BASE_ADDR)];
111 	p4d = p4d_alloc(&init_mm, pgd, ESPFIX_BASE_ADDR);
112 	p4d_populate(&init_mm, p4d, espfix_pud_page);
113 
114 	/* Randomize the locations */
115 	init_espfix_random();
116 
117 	/* The rest is the same as for any other processor */
118 	init_espfix_ap(0);
119 }
120 
121 void init_espfix_ap(int cpu)
122 {
123 	unsigned int page;
124 	unsigned long addr;
125 	pud_t pud, *pud_p;
126 	pmd_t pmd, *pmd_p;
127 	pte_t pte, *pte_p;
128 	int n, node;
129 	void *stack_page;
130 	pteval_t ptemask;
131 
132 	/* We only have to do this once... */
133 	if (likely(per_cpu(espfix_stack, cpu)))
134 		return;		/* Already initialized */
135 
136 	addr = espfix_base_addr(cpu);
137 	page = cpu/ESPFIX_STACKS_PER_PAGE;
138 
139 	/* Did another CPU already set this up? */
140 	stack_page = READ_ONCE(espfix_pages[page]);
141 	if (likely(stack_page))
142 		goto done;
143 
144 	mutex_lock(&espfix_init_mutex);
145 
146 	/* Did we race on the lock? */
147 	stack_page = READ_ONCE(espfix_pages[page]);
148 	if (stack_page)
149 		goto unlock_done;
150 
151 	node = cpu_to_node(cpu);
152 	ptemask = __supported_pte_mask;
153 
154 	pud_p = &espfix_pud_page[pud_index(addr)];
155 	pud = *pud_p;
156 	if (!pud_present(pud)) {
157 		struct page *page = alloc_pages_node(node, PGALLOC_GFP, 0);
158 
159 		pmd_p = (pmd_t *)page_address(page);
160 		pud = __pud(__pa(pmd_p) | (PGTABLE_PROT & ptemask));
161 		paravirt_alloc_pmd(&init_mm, __pa(pmd_p) >> PAGE_SHIFT);
162 		for (n = 0; n < ESPFIX_PUD_CLONES; n++)
163 			set_pud(&pud_p[n], pud);
164 	}
165 
166 	pmd_p = pmd_offset(&pud, addr);
167 	pmd = *pmd_p;
168 	if (!pmd_present(pmd)) {
169 		struct page *page = alloc_pages_node(node, PGALLOC_GFP, 0);
170 
171 		pte_p = (pte_t *)page_address(page);
172 		pmd = __pmd(__pa(pte_p) | (PGTABLE_PROT & ptemask));
173 		paravirt_alloc_pte(&init_mm, __pa(pte_p) >> PAGE_SHIFT);
174 		for (n = 0; n < ESPFIX_PMD_CLONES; n++)
175 			set_pmd(&pmd_p[n], pmd);
176 	}
177 
178 	pte_p = pte_offset_kernel(&pmd, addr);
179 	stack_page = page_address(alloc_pages_node(node, GFP_KERNEL, 0));
180 	/*
181 	 * __PAGE_KERNEL_* includes _PAGE_GLOBAL, which we want since
182 	 * this is mapped to userspace.
183 	 */
184 	pte = __pte(__pa(stack_page) | ((__PAGE_KERNEL_RO | _PAGE_ENC) & ptemask));
185 	for (n = 0; n < ESPFIX_PTE_CLONES; n++)
186 		set_pte(&pte_p[n*PTE_STRIDE], pte);
187 
188 	/* Job is done for this CPU and any CPU which shares this page */
189 	WRITE_ONCE(espfix_pages[page], stack_page);
190 
191 unlock_done:
192 	mutex_unlock(&espfix_init_mutex);
193 done:
194 	per_cpu(espfix_stack, cpu) = addr;
195 	per_cpu(espfix_waddr, cpu) = (unsigned long)stack_page
196 				      + (addr & ~PAGE_MASK);
197 }
198