xref: /openbmc/linux/arch/arm64/kernel/alternative.c (revision 0e407915)
1 // SPDX-License-Identifier: GPL-2.0-only
2 /*
3  * alternative runtime patching
4  * inspired by the x86 version
5  *
6  * Copyright (C) 2014 ARM Ltd.
7  */
8 
9 #define pr_fmt(fmt) "alternatives: " fmt
10 
11 #include <linux/init.h>
12 #include <linux/cpu.h>
13 #include <asm/cacheflush.h>
14 #include <asm/alternative.h>
15 #include <asm/cpufeature.h>
16 #include <asm/insn.h>
17 #include <asm/sections.h>
18 #include <linux/stop_machine.h>
19 
20 #define __ALT_PTR(a, f)		((void *)&(a)->f + (a)->f)
21 #define ALT_ORIG_PTR(a)		__ALT_PTR(a, orig_offset)
22 #define ALT_REPL_PTR(a)		__ALT_PTR(a, alt_offset)
23 
24 /* Volatile, as we may be patching the guts of READ_ONCE() */
25 static volatile int all_alternatives_applied;
26 
27 static DECLARE_BITMAP(applied_alternatives, ARM64_NCAPS);
28 
29 struct alt_region {
30 	struct alt_instr *begin;
31 	struct alt_instr *end;
32 };
33 
34 bool alternative_is_applied(u16 cpufeature)
35 {
36 	if (WARN_ON(cpufeature >= ARM64_NCAPS))
37 		return false;
38 
39 	return test_bit(cpufeature, applied_alternatives);
40 }
41 
42 /*
43  * Check if the target PC is within an alternative block.
44  */
45 static __always_inline bool branch_insn_requires_update(struct alt_instr *alt, unsigned long pc)
46 {
47 	unsigned long replptr = (unsigned long)ALT_REPL_PTR(alt);
48 	return !(pc >= replptr && pc <= (replptr + alt->alt_len));
49 }
50 
51 #define align_down(x, a)	((unsigned long)(x) & ~(((unsigned long)(a)) - 1))
52 
53 static __always_inline u32 get_alt_insn(struct alt_instr *alt, __le32 *insnptr, __le32 *altinsnptr)
54 {
55 	u32 insn;
56 
57 	insn = le32_to_cpu(*altinsnptr);
58 
59 	if (aarch64_insn_is_branch_imm(insn)) {
60 		s32 offset = aarch64_get_branch_offset(insn);
61 		unsigned long target;
62 
63 		target = (unsigned long)altinsnptr + offset;
64 
65 		/*
66 		 * If we're branching inside the alternate sequence,
67 		 * do not rewrite the instruction, as it is already
68 		 * correct. Otherwise, generate the new instruction.
69 		 */
70 		if (branch_insn_requires_update(alt, target)) {
71 			offset = target - (unsigned long)insnptr;
72 			insn = aarch64_set_branch_offset(insn, offset);
73 		}
74 	} else if (aarch64_insn_is_adrp(insn)) {
75 		s32 orig_offset, new_offset;
76 		unsigned long target;
77 
78 		/*
79 		 * If we're replacing an adrp instruction, which uses PC-relative
80 		 * immediate addressing, adjust the offset to reflect the new
81 		 * PC. adrp operates on 4K aligned addresses.
82 		 */
83 		orig_offset  = aarch64_insn_adrp_get_offset(insn);
84 		target = align_down(altinsnptr, SZ_4K) + orig_offset;
85 		new_offset = target - align_down(insnptr, SZ_4K);
86 		insn = aarch64_insn_adrp_set_offset(insn, new_offset);
87 	} else if (aarch64_insn_uses_literal(insn)) {
88 		/*
89 		 * Disallow patching unhandled instructions using PC relative
90 		 * literal addresses
91 		 */
92 		BUG();
93 	}
94 
95 	return insn;
96 }
97 
98 static noinstr void patch_alternative(struct alt_instr *alt,
99 			      __le32 *origptr, __le32 *updptr, int nr_inst)
100 {
101 	__le32 *replptr;
102 	int i;
103 
104 	replptr = ALT_REPL_PTR(alt);
105 	for (i = 0; i < nr_inst; i++) {
106 		u32 insn;
107 
108 		insn = get_alt_insn(alt, origptr + i, replptr + i);
109 		updptr[i] = cpu_to_le32(insn);
110 	}
111 }
112 
113 /*
114  * We provide our own, private D-cache cleaning function so that we don't
115  * accidentally call into the cache.S code, which is patched by us at
116  * runtime.
117  */
118 static void clean_dcache_range_nopatch(u64 start, u64 end)
119 {
120 	u64 cur, d_size, ctr_el0;
121 
122 	ctr_el0 = read_sanitised_ftr_reg(SYS_CTR_EL0);
123 	d_size = 4 << cpuid_feature_extract_unsigned_field(ctr_el0,
124 							   CTR_EL0_DminLine_SHIFT);
125 	cur = start & ~(d_size - 1);
126 	do {
127 		/*
128 		 * We must clean+invalidate to the PoC in order to avoid
129 		 * Cortex-A53 errata 826319, 827319, 824069 and 819472
130 		 * (this corresponds to ARM64_WORKAROUND_CLEAN_CACHE)
131 		 */
132 		asm volatile("dc civac, %0" : : "r" (cur) : "memory");
133 	} while (cur += d_size, cur < end);
134 }
135 
136 static void __apply_alternatives(struct alt_region *region, bool is_module,
137 				 unsigned long *feature_mask)
138 {
139 	struct alt_instr *alt;
140 	__le32 *origptr, *updptr;
141 	alternative_cb_t alt_cb;
142 
143 	for (alt = region->begin; alt < region->end; alt++) {
144 		int nr_inst;
145 
146 		if (!test_bit(alt->cpufeature, feature_mask))
147 			continue;
148 
149 		/* Use ARM64_CB_PATCH as an unconditional patch */
150 		if (alt->cpufeature < ARM64_CB_PATCH &&
151 		    !cpus_have_cap(alt->cpufeature))
152 			continue;
153 
154 		if (alt->cpufeature == ARM64_CB_PATCH)
155 			BUG_ON(alt->alt_len != 0);
156 		else
157 			BUG_ON(alt->alt_len != alt->orig_len);
158 
159 		pr_info_once("patching kernel code\n");
160 
161 		origptr = ALT_ORIG_PTR(alt);
162 		updptr = is_module ? origptr : lm_alias(origptr);
163 		nr_inst = alt->orig_len / AARCH64_INSN_SIZE;
164 
165 		if (alt->cpufeature < ARM64_CB_PATCH)
166 			alt_cb = patch_alternative;
167 		else
168 			alt_cb  = ALT_REPL_PTR(alt);
169 
170 		alt_cb(alt, origptr, updptr, nr_inst);
171 
172 		if (!is_module) {
173 			clean_dcache_range_nopatch((u64)origptr,
174 						   (u64)(origptr + nr_inst));
175 		}
176 	}
177 
178 	/*
179 	 * The core module code takes care of cache maintenance in
180 	 * flush_module_icache().
181 	 */
182 	if (!is_module) {
183 		dsb(ish);
184 		icache_inval_all_pou();
185 		isb();
186 
187 		/* Ignore ARM64_CB bit from feature mask */
188 		bitmap_or(applied_alternatives, applied_alternatives,
189 			  feature_mask, ARM64_NCAPS);
190 		bitmap_and(applied_alternatives, applied_alternatives,
191 			   cpu_hwcaps, ARM64_NCAPS);
192 	}
193 }
194 
195 /*
196  * We might be patching the stop_machine state machine, so implement a
197  * really simple polling protocol here.
198  */
199 static int __apply_alternatives_multi_stop(void *unused)
200 {
201 	struct alt_region region = {
202 		.begin	= (struct alt_instr *)__alt_instructions,
203 		.end	= (struct alt_instr *)__alt_instructions_end,
204 	};
205 
206 	/* We always have a CPU 0 at this point (__init) */
207 	if (smp_processor_id()) {
208 		while (!all_alternatives_applied)
209 			cpu_relax();
210 		isb();
211 	} else {
212 		DECLARE_BITMAP(remaining_capabilities, ARM64_NPATCHABLE);
213 
214 		bitmap_complement(remaining_capabilities, boot_capabilities,
215 				  ARM64_NPATCHABLE);
216 
217 		BUG_ON(all_alternatives_applied);
218 		__apply_alternatives(&region, false, remaining_capabilities);
219 		/* Barriers provided by the cache flushing */
220 		all_alternatives_applied = 1;
221 	}
222 
223 	return 0;
224 }
225 
226 void __init apply_alternatives_all(void)
227 {
228 	/* better not try code patching on a live SMP system */
229 	stop_machine(__apply_alternatives_multi_stop, NULL, cpu_online_mask);
230 }
231 
232 /*
233  * This is called very early in the boot process (directly after we run
234  * a feature detect on the boot CPU). No need to worry about other CPUs
235  * here.
236  */
237 void __init apply_boot_alternatives(void)
238 {
239 	struct alt_region region = {
240 		.begin	= (struct alt_instr *)__alt_instructions,
241 		.end	= (struct alt_instr *)__alt_instructions_end,
242 	};
243 
244 	/* If called on non-boot cpu things could go wrong */
245 	WARN_ON(smp_processor_id() != 0);
246 
247 	__apply_alternatives(&region, false, &boot_capabilities[0]);
248 }
249 
250 #ifdef CONFIG_MODULES
251 void apply_alternatives_module(void *start, size_t length)
252 {
253 	struct alt_region region = {
254 		.begin	= start,
255 		.end	= start + length,
256 	};
257 	DECLARE_BITMAP(all_capabilities, ARM64_NPATCHABLE);
258 
259 	bitmap_fill(all_capabilities, ARM64_NPATCHABLE);
260 
261 	__apply_alternatives(&region, true, &all_capabilities[0]);
262 }
263 #endif
264