xref: /openbmc/linux/arch/powerpc/lib/code-patching.c (revision 980411a4)
1 // SPDX-License-Identifier: GPL-2.0-or-later
2 /*
3  *  Copyright 2008 Michael Ellerman, IBM Corporation.
4  */
5 
6 #include <linux/kprobes.h>
7 #include <linux/mmu_context.h>
8 #include <linux/random.h>
9 #include <linux/vmalloc.h>
10 #include <linux/init.h>
11 #include <linux/cpuhotplug.h>
12 #include <linux/uaccess.h>
13 #include <linux/jump_label.h>
14 
15 #include <asm/debug.h>
16 #include <asm/pgalloc.h>
17 #include <asm/tlb.h>
18 #include <asm/tlbflush.h>
19 #include <asm/page.h>
20 #include <asm/code-patching.h>
21 #include <asm/inst.h>
22 
__patch_instruction(u32 * exec_addr,ppc_inst_t instr,u32 * patch_addr)23 static int __patch_instruction(u32 *exec_addr, ppc_inst_t instr, u32 *patch_addr)
24 {
25 	if (!ppc_inst_prefixed(instr)) {
26 		u32 val = ppc_inst_val(instr);
27 
28 		__put_kernel_nofault(patch_addr, &val, u32, failed);
29 	} else {
30 		u64 val = ppc_inst_as_ulong(instr);
31 
32 		__put_kernel_nofault(patch_addr, &val, u64, failed);
33 	}
34 
35 	asm ("dcbst 0, %0; sync; icbi 0,%1; sync; isync" :: "r" (patch_addr),
36 							    "r" (exec_addr));
37 
38 	return 0;
39 
40 failed:
41 	return -EPERM;
42 }
43 
raw_patch_instruction(u32 * addr,ppc_inst_t instr)44 int raw_patch_instruction(u32 *addr, ppc_inst_t instr)
45 {
46 	return __patch_instruction(addr, instr, addr);
47 }
48 
49 struct patch_context {
50 	union {
51 		struct vm_struct *area;
52 		struct mm_struct *mm;
53 	};
54 	unsigned long addr;
55 	pte_t *pte;
56 };
57 
58 static DEFINE_PER_CPU(struct patch_context, cpu_patching_context);
59 
60 static int map_patch_area(void *addr, unsigned long text_poke_addr);
61 static void unmap_patch_area(unsigned long addr);
62 
mm_patch_enabled(void)63 static bool mm_patch_enabled(void)
64 {
65 	return IS_ENABLED(CONFIG_SMP) && radix_enabled();
66 }
67 
68 /*
69  * The following applies for Radix MMU. Hash MMU has different requirements,
70  * and so is not supported.
71  *
72  * Changing mm requires context synchronising instructions on both sides of
73  * the context switch, as well as a hwsync between the last instruction for
74  * which the address of an associated storage access was translated using
75  * the current context.
76  *
77  * switch_mm_irqs_off() performs an isync after the context switch. It is
78  * the responsibility of the caller to perform the CSI and hwsync before
79  * starting/stopping the temp mm.
80  */
start_using_temp_mm(struct mm_struct * temp_mm)81 static struct mm_struct *start_using_temp_mm(struct mm_struct *temp_mm)
82 {
83 	struct mm_struct *orig_mm = current->active_mm;
84 
85 	lockdep_assert_irqs_disabled();
86 	switch_mm_irqs_off(orig_mm, temp_mm, current);
87 
88 	WARN_ON(!mm_is_thread_local(temp_mm));
89 
90 	suspend_breakpoints();
91 	return orig_mm;
92 }
93 
stop_using_temp_mm(struct mm_struct * temp_mm,struct mm_struct * orig_mm)94 static void stop_using_temp_mm(struct mm_struct *temp_mm,
95 			       struct mm_struct *orig_mm)
96 {
97 	lockdep_assert_irqs_disabled();
98 	switch_mm_irqs_off(temp_mm, orig_mm, current);
99 	restore_breakpoints();
100 }
101 
text_area_cpu_up(unsigned int cpu)102 static int text_area_cpu_up(unsigned int cpu)
103 {
104 	struct vm_struct *area;
105 	unsigned long addr;
106 	int err;
107 
108 	area = get_vm_area(PAGE_SIZE, VM_ALLOC);
109 	if (!area) {
110 		WARN_ONCE(1, "Failed to create text area for cpu %d\n",
111 			cpu);
112 		return -1;
113 	}
114 
115 	// Map/unmap the area to ensure all page tables are pre-allocated
116 	addr = (unsigned long)area->addr;
117 	err = map_patch_area(empty_zero_page, addr);
118 	if (err)
119 		return err;
120 
121 	unmap_patch_area(addr);
122 
123 	this_cpu_write(cpu_patching_context.area, area);
124 	this_cpu_write(cpu_patching_context.addr, addr);
125 	this_cpu_write(cpu_patching_context.pte, virt_to_kpte(addr));
126 
127 	return 0;
128 }
129 
text_area_cpu_down(unsigned int cpu)130 static int text_area_cpu_down(unsigned int cpu)
131 {
132 	free_vm_area(this_cpu_read(cpu_patching_context.area));
133 	this_cpu_write(cpu_patching_context.area, NULL);
134 	this_cpu_write(cpu_patching_context.addr, 0);
135 	this_cpu_write(cpu_patching_context.pte, NULL);
136 	return 0;
137 }
138 
put_patching_mm(struct mm_struct * mm,unsigned long patching_addr)139 static void put_patching_mm(struct mm_struct *mm, unsigned long patching_addr)
140 {
141 	struct mmu_gather tlb;
142 
143 	tlb_gather_mmu(&tlb, mm);
144 	free_pgd_range(&tlb, patching_addr, patching_addr + PAGE_SIZE, 0, 0);
145 	mmput(mm);
146 }
147 
text_area_cpu_up_mm(unsigned int cpu)148 static int text_area_cpu_up_mm(unsigned int cpu)
149 {
150 	struct mm_struct *mm;
151 	unsigned long addr;
152 	pte_t *pte;
153 	spinlock_t *ptl;
154 
155 	mm = mm_alloc();
156 	if (WARN_ON(!mm))
157 		goto fail_no_mm;
158 
159 	/*
160 	 * Choose a random page-aligned address from the interval
161 	 * [PAGE_SIZE .. DEFAULT_MAP_WINDOW - PAGE_SIZE].
162 	 * The lower address bound is PAGE_SIZE to avoid the zero-page.
163 	 */
164 	addr = (1 + (get_random_long() % (DEFAULT_MAP_WINDOW / PAGE_SIZE - 2))) << PAGE_SHIFT;
165 
166 	/*
167 	 * PTE allocation uses GFP_KERNEL which means we need to
168 	 * pre-allocate the PTE here because we cannot do the
169 	 * allocation during patching when IRQs are disabled.
170 	 *
171 	 * Using get_locked_pte() to avoid open coding, the lock
172 	 * is unnecessary.
173 	 */
174 	pte = get_locked_pte(mm, addr, &ptl);
175 	if (!pte)
176 		goto fail_no_pte;
177 	pte_unmap_unlock(pte, ptl);
178 
179 	this_cpu_write(cpu_patching_context.mm, mm);
180 	this_cpu_write(cpu_patching_context.addr, addr);
181 
182 	return 0;
183 
184 fail_no_pte:
185 	put_patching_mm(mm, addr);
186 fail_no_mm:
187 	return -ENOMEM;
188 }
189 
text_area_cpu_down_mm(unsigned int cpu)190 static int text_area_cpu_down_mm(unsigned int cpu)
191 {
192 	put_patching_mm(this_cpu_read(cpu_patching_context.mm),
193 			this_cpu_read(cpu_patching_context.addr));
194 
195 	this_cpu_write(cpu_patching_context.mm, NULL);
196 	this_cpu_write(cpu_patching_context.addr, 0);
197 
198 	return 0;
199 }
200 
201 static __ro_after_init DEFINE_STATIC_KEY_FALSE(poking_init_done);
202 
poking_init(void)203 void __init poking_init(void)
204 {
205 	int ret;
206 
207 	if (!IS_ENABLED(CONFIG_STRICT_KERNEL_RWX))
208 		return;
209 
210 	if (mm_patch_enabled())
211 		ret = cpuhp_setup_state(CPUHP_AP_ONLINE_DYN,
212 					"powerpc/text_poke_mm:online",
213 					text_area_cpu_up_mm,
214 					text_area_cpu_down_mm);
215 	else
216 		ret = cpuhp_setup_state(CPUHP_AP_ONLINE_DYN,
217 					"powerpc/text_poke:online",
218 					text_area_cpu_up,
219 					text_area_cpu_down);
220 
221 	/* cpuhp_setup_state returns >= 0 on success */
222 	if (WARN_ON(ret < 0))
223 		return;
224 
225 	static_branch_enable(&poking_init_done);
226 }
227 
get_patch_pfn(void * addr)228 static unsigned long get_patch_pfn(void *addr)
229 {
230 	if (IS_ENABLED(CONFIG_MODULES) && is_vmalloc_or_module_addr(addr))
231 		return vmalloc_to_pfn(addr);
232 	else
233 		return __pa_symbol(addr) >> PAGE_SHIFT;
234 }
235 
236 /*
237  * This can be called for kernel text or a module.
238  */
map_patch_area(void * addr,unsigned long text_poke_addr)239 static int map_patch_area(void *addr, unsigned long text_poke_addr)
240 {
241 	unsigned long pfn = get_patch_pfn(addr);
242 
243 	return map_kernel_page(text_poke_addr, (pfn << PAGE_SHIFT), PAGE_KERNEL);
244 }
245 
unmap_patch_area(unsigned long addr)246 static void unmap_patch_area(unsigned long addr)
247 {
248 	pte_t *ptep;
249 	pmd_t *pmdp;
250 	pud_t *pudp;
251 	p4d_t *p4dp;
252 	pgd_t *pgdp;
253 
254 	pgdp = pgd_offset_k(addr);
255 	if (WARN_ON(pgd_none(*pgdp)))
256 		return;
257 
258 	p4dp = p4d_offset(pgdp, addr);
259 	if (WARN_ON(p4d_none(*p4dp)))
260 		return;
261 
262 	pudp = pud_offset(p4dp, addr);
263 	if (WARN_ON(pud_none(*pudp)))
264 		return;
265 
266 	pmdp = pmd_offset(pudp, addr);
267 	if (WARN_ON(pmd_none(*pmdp)))
268 		return;
269 
270 	ptep = pte_offset_kernel(pmdp, addr);
271 	if (WARN_ON(pte_none(*ptep)))
272 		return;
273 
274 	/*
275 	 * In hash, pte_clear flushes the tlb, in radix, we have to
276 	 */
277 	pte_clear(&init_mm, addr, ptep);
278 	flush_tlb_kernel_range(addr, addr + PAGE_SIZE);
279 }
280 
__do_patch_instruction_mm(u32 * addr,ppc_inst_t instr)281 static int __do_patch_instruction_mm(u32 *addr, ppc_inst_t instr)
282 {
283 	int err;
284 	u32 *patch_addr;
285 	unsigned long text_poke_addr;
286 	pte_t *pte;
287 	unsigned long pfn = get_patch_pfn(addr);
288 	struct mm_struct *patching_mm;
289 	struct mm_struct *orig_mm;
290 	spinlock_t *ptl;
291 
292 	patching_mm = __this_cpu_read(cpu_patching_context.mm);
293 	text_poke_addr = __this_cpu_read(cpu_patching_context.addr);
294 	patch_addr = (u32 *)(text_poke_addr + offset_in_page(addr));
295 
296 	pte = get_locked_pte(patching_mm, text_poke_addr, &ptl);
297 	if (!pte)
298 		return -ENOMEM;
299 
300 	__set_pte_at(patching_mm, text_poke_addr, pte, pfn_pte(pfn, PAGE_KERNEL), 0);
301 
302 	/* order PTE update before use, also serves as the hwsync */
303 	asm volatile("ptesync": : :"memory");
304 
305 	/* order context switch after arbitrary prior code */
306 	isync();
307 
308 	orig_mm = start_using_temp_mm(patching_mm);
309 
310 	err = __patch_instruction(addr, instr, patch_addr);
311 
312 	/* hwsync performed by __patch_instruction (sync) if successful */
313 	if (err)
314 		mb();  /* sync */
315 
316 	/* context synchronisation performed by __patch_instruction (isync or exception) */
317 	stop_using_temp_mm(patching_mm, orig_mm);
318 
319 	pte_clear(patching_mm, text_poke_addr, pte);
320 	/*
321 	 * ptesync to order PTE update before TLB invalidation done
322 	 * by radix__local_flush_tlb_page_psize (in _tlbiel_va)
323 	 */
324 	local_flush_tlb_page_psize(patching_mm, text_poke_addr, mmu_virtual_psize);
325 
326 	pte_unmap_unlock(pte, ptl);
327 
328 	return err;
329 }
330 
__do_patch_instruction(u32 * addr,ppc_inst_t instr)331 static int __do_patch_instruction(u32 *addr, ppc_inst_t instr)
332 {
333 	int err;
334 	u32 *patch_addr;
335 	unsigned long text_poke_addr;
336 	pte_t *pte;
337 	unsigned long pfn = get_patch_pfn(addr);
338 
339 	text_poke_addr = (unsigned long)__this_cpu_read(cpu_patching_context.addr) & PAGE_MASK;
340 	patch_addr = (u32 *)(text_poke_addr + offset_in_page(addr));
341 
342 	pte = __this_cpu_read(cpu_patching_context.pte);
343 	__set_pte_at(&init_mm, text_poke_addr, pte, pfn_pte(pfn, PAGE_KERNEL), 0);
344 	/* See ptesync comment in radix__set_pte_at() */
345 	if (radix_enabled())
346 		asm volatile("ptesync": : :"memory");
347 
348 	err = __patch_instruction(addr, instr, patch_addr);
349 
350 	pte_clear(&init_mm, text_poke_addr, pte);
351 	flush_tlb_kernel_range(text_poke_addr, text_poke_addr + PAGE_SIZE);
352 
353 	return err;
354 }
355 
patch_instruction(u32 * addr,ppc_inst_t instr)356 int patch_instruction(u32 *addr, ppc_inst_t instr)
357 {
358 	int err;
359 	unsigned long flags;
360 
361 	/*
362 	 * During early early boot patch_instruction is called
363 	 * when text_poke_area is not ready, but we still need
364 	 * to allow patching. We just do the plain old patching
365 	 */
366 	if (!IS_ENABLED(CONFIG_STRICT_KERNEL_RWX) ||
367 	    !static_branch_likely(&poking_init_done))
368 		return raw_patch_instruction(addr, instr);
369 
370 	local_irq_save(flags);
371 	if (mm_patch_enabled())
372 		err = __do_patch_instruction_mm(addr, instr);
373 	else
374 		err = __do_patch_instruction(addr, instr);
375 	local_irq_restore(flags);
376 
377 	return err;
378 }
379 NOKPROBE_SYMBOL(patch_instruction);
380 
patch_branch(u32 * addr,unsigned long target,int flags)381 int patch_branch(u32 *addr, unsigned long target, int flags)
382 {
383 	ppc_inst_t instr;
384 
385 	if (create_branch(&instr, addr, target, flags))
386 		return -ERANGE;
387 
388 	return patch_instruction(addr, instr);
389 }
390 
391 /*
392  * Helper to check if a given instruction is a conditional branch
393  * Derived from the conditional checks in analyse_instr()
394  */
is_conditional_branch(ppc_inst_t instr)395 bool is_conditional_branch(ppc_inst_t instr)
396 {
397 	unsigned int opcode = ppc_inst_primary_opcode(instr);
398 
399 	if (opcode == 16)       /* bc, bca, bcl, bcla */
400 		return true;
401 	if (opcode == 19) {
402 		switch ((ppc_inst_val(instr) >> 1) & 0x3ff) {
403 		case 16:        /* bclr, bclrl */
404 		case 528:       /* bcctr, bcctrl */
405 		case 560:       /* bctar, bctarl */
406 			return true;
407 		}
408 	}
409 	return false;
410 }
411 NOKPROBE_SYMBOL(is_conditional_branch);
412 
create_cond_branch(ppc_inst_t * instr,const u32 * addr,unsigned long target,int flags)413 int create_cond_branch(ppc_inst_t *instr, const u32 *addr,
414 		       unsigned long target, int flags)
415 {
416 	long offset;
417 
418 	offset = target;
419 	if (! (flags & BRANCH_ABSOLUTE))
420 		offset = offset - (unsigned long)addr;
421 
422 	/* Check we can represent the target in the instruction format */
423 	if (!is_offset_in_cond_branch_range(offset))
424 		return 1;
425 
426 	/* Mask out the flags and target, so they don't step on each other. */
427 	*instr = ppc_inst(0x40000000 | (flags & 0x3FF0003) | (offset & 0xFFFC));
428 
429 	return 0;
430 }
431 
instr_is_relative_branch(ppc_inst_t instr)432 int instr_is_relative_branch(ppc_inst_t instr)
433 {
434 	if (ppc_inst_val(instr) & BRANCH_ABSOLUTE)
435 		return 0;
436 
437 	return instr_is_branch_iform(instr) || instr_is_branch_bform(instr);
438 }
439 
instr_is_relative_link_branch(ppc_inst_t instr)440 int instr_is_relative_link_branch(ppc_inst_t instr)
441 {
442 	return instr_is_relative_branch(instr) && (ppc_inst_val(instr) & BRANCH_SET_LINK);
443 }
444 
branch_iform_target(const u32 * instr)445 static unsigned long branch_iform_target(const u32 *instr)
446 {
447 	signed long imm;
448 
449 	imm = ppc_inst_val(ppc_inst_read(instr)) & 0x3FFFFFC;
450 
451 	/* If the top bit of the immediate value is set this is negative */
452 	if (imm & 0x2000000)
453 		imm -= 0x4000000;
454 
455 	if ((ppc_inst_val(ppc_inst_read(instr)) & BRANCH_ABSOLUTE) == 0)
456 		imm += (unsigned long)instr;
457 
458 	return (unsigned long)imm;
459 }
460 
branch_bform_target(const u32 * instr)461 static unsigned long branch_bform_target(const u32 *instr)
462 {
463 	signed long imm;
464 
465 	imm = ppc_inst_val(ppc_inst_read(instr)) & 0xFFFC;
466 
467 	/* If the top bit of the immediate value is set this is negative */
468 	if (imm & 0x8000)
469 		imm -= 0x10000;
470 
471 	if ((ppc_inst_val(ppc_inst_read(instr)) & BRANCH_ABSOLUTE) == 0)
472 		imm += (unsigned long)instr;
473 
474 	return (unsigned long)imm;
475 }
476 
branch_target(const u32 * instr)477 unsigned long branch_target(const u32 *instr)
478 {
479 	if (instr_is_branch_iform(ppc_inst_read(instr)))
480 		return branch_iform_target(instr);
481 	else if (instr_is_branch_bform(ppc_inst_read(instr)))
482 		return branch_bform_target(instr);
483 
484 	return 0;
485 }
486 
translate_branch(ppc_inst_t * instr,const u32 * dest,const u32 * src)487 int translate_branch(ppc_inst_t *instr, const u32 *dest, const u32 *src)
488 {
489 	unsigned long target;
490 	target = branch_target(src);
491 
492 	if (instr_is_branch_iform(ppc_inst_read(src)))
493 		return create_branch(instr, dest, target,
494 				     ppc_inst_val(ppc_inst_read(src)));
495 	else if (instr_is_branch_bform(ppc_inst_read(src)))
496 		return create_cond_branch(instr, dest, target,
497 					  ppc_inst_val(ppc_inst_read(src)));
498 
499 	return 1;
500 }
501