Linux kernel mirror (for testing)
git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git
kernel
os
linux
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
23static 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
40failed:
41 mb(); /* sync */
42 return -EPERM;
43}
44
45int raw_patch_instruction(u32 *addr, ppc_inst_t instr)
46{
47 return __patch_instruction(addr, instr, addr);
48}
49
50struct patch_context {
51 union {
52 struct vm_struct *area;
53 struct mm_struct *mm;
54 };
55 unsigned long addr;
56 pte_t *pte;
57};
58
59static DEFINE_PER_CPU(struct patch_context, cpu_patching_context);
60
61static int map_patch_area(void *addr, unsigned long text_poke_addr);
62static void unmap_patch_area(unsigned long addr);
63
64static bool mm_patch_enabled(void)
65{
66 return IS_ENABLED(CONFIG_SMP) && radix_enabled();
67}
68
69/*
70 * The following applies for Radix MMU. Hash MMU has different requirements,
71 * and so is not supported.
72 *
73 * Changing mm requires context synchronising instructions on both sides of
74 * the context switch, as well as a hwsync between the last instruction for
75 * which the address of an associated storage access was translated using
76 * the current context.
77 *
78 * switch_mm_irqs_off() performs an isync after the context switch. It is
79 * the responsibility of the caller to perform the CSI and hwsync before
80 * starting/stopping the temp mm.
81 */
82static struct mm_struct *start_using_temp_mm(struct mm_struct *temp_mm)
83{
84 struct mm_struct *orig_mm = current->active_mm;
85
86 lockdep_assert_irqs_disabled();
87 switch_mm_irqs_off(orig_mm, temp_mm, current);
88
89 WARN_ON(!mm_is_thread_local(temp_mm));
90
91 suspend_breakpoints();
92 return orig_mm;
93}
94
95static void stop_using_temp_mm(struct mm_struct *temp_mm,
96 struct mm_struct *orig_mm)
97{
98 lockdep_assert_irqs_disabled();
99 switch_mm_irqs_off(temp_mm, orig_mm, current);
100 restore_breakpoints();
101}
102
103static int text_area_cpu_up(unsigned int cpu)
104{
105 struct vm_struct *area;
106 unsigned long addr;
107 int err;
108
109 area = get_vm_area(PAGE_SIZE, VM_ALLOC);
110 if (!area) {
111 WARN_ONCE(1, "Failed to create text area for cpu %d\n",
112 cpu);
113 return -1;
114 }
115
116 // Map/unmap the area to ensure all page tables are pre-allocated
117 addr = (unsigned long)area->addr;
118 err = map_patch_area(empty_zero_page, addr);
119 if (err)
120 return err;
121
122 unmap_patch_area(addr);
123
124 this_cpu_write(cpu_patching_context.area, area);
125 this_cpu_write(cpu_patching_context.addr, addr);
126 this_cpu_write(cpu_patching_context.pte, virt_to_kpte(addr));
127
128 return 0;
129}
130
131static int text_area_cpu_down(unsigned int cpu)
132{
133 free_vm_area(this_cpu_read(cpu_patching_context.area));
134 this_cpu_write(cpu_patching_context.area, NULL);
135 this_cpu_write(cpu_patching_context.addr, 0);
136 this_cpu_write(cpu_patching_context.pte, NULL);
137 return 0;
138}
139
140static void put_patching_mm(struct mm_struct *mm, unsigned long patching_addr)
141{
142 struct mmu_gather tlb;
143
144 tlb_gather_mmu(&tlb, mm);
145 free_pgd_range(&tlb, patching_addr, patching_addr + PAGE_SIZE, 0, 0);
146 mmput(mm);
147}
148
149static int text_area_cpu_up_mm(unsigned int cpu)
150{
151 struct mm_struct *mm;
152 unsigned long addr;
153 pte_t *pte;
154 spinlock_t *ptl;
155
156 mm = mm_alloc();
157 if (WARN_ON(!mm))
158 goto fail_no_mm;
159
160 /*
161 * Choose a random page-aligned address from the interval
162 * [PAGE_SIZE .. DEFAULT_MAP_WINDOW - PAGE_SIZE].
163 * The lower address bound is PAGE_SIZE to avoid the zero-page.
164 */
165 addr = (1 + (get_random_long() % (DEFAULT_MAP_WINDOW / PAGE_SIZE - 2))) << PAGE_SHIFT;
166
167 /*
168 * PTE allocation uses GFP_KERNEL which means we need to
169 * pre-allocate the PTE here because we cannot do the
170 * allocation during patching when IRQs are disabled.
171 *
172 * Using get_locked_pte() to avoid open coding, the lock
173 * is unnecessary.
174 */
175 pte = get_locked_pte(mm, addr, &ptl);
176 if (!pte)
177 goto fail_no_pte;
178 pte_unmap_unlock(pte, ptl);
179
180 this_cpu_write(cpu_patching_context.mm, mm);
181 this_cpu_write(cpu_patching_context.addr, addr);
182
183 return 0;
184
185fail_no_pte:
186 put_patching_mm(mm, addr);
187fail_no_mm:
188 return -ENOMEM;
189}
190
191static int text_area_cpu_down_mm(unsigned int cpu)
192{
193 put_patching_mm(this_cpu_read(cpu_patching_context.mm),
194 this_cpu_read(cpu_patching_context.addr));
195
196 this_cpu_write(cpu_patching_context.mm, NULL);
197 this_cpu_write(cpu_patching_context.addr, 0);
198
199 return 0;
200}
201
202static __ro_after_init DEFINE_STATIC_KEY_FALSE(poking_init_done);
203
204void __init poking_init(void)
205{
206 int ret;
207
208 if (mm_patch_enabled())
209 ret = cpuhp_setup_state(CPUHP_AP_ONLINE_DYN,
210 "powerpc/text_poke_mm:online",
211 text_area_cpu_up_mm,
212 text_area_cpu_down_mm);
213 else
214 ret = cpuhp_setup_state(CPUHP_AP_ONLINE_DYN,
215 "powerpc/text_poke:online",
216 text_area_cpu_up,
217 text_area_cpu_down);
218
219 /* cpuhp_setup_state returns >= 0 on success */
220 if (WARN_ON(ret < 0))
221 return;
222
223 static_branch_enable(&poking_init_done);
224}
225
226static unsigned long get_patch_pfn(void *addr)
227{
228 if (IS_ENABLED(CONFIG_EXECMEM) && is_vmalloc_or_module_addr(addr))
229 return vmalloc_to_pfn(addr);
230 else
231 return __pa_symbol(addr) >> PAGE_SHIFT;
232}
233
234/*
235 * This can be called for kernel text or a module.
236 */
237static int map_patch_area(void *addr, unsigned long text_poke_addr)
238{
239 unsigned long pfn = get_patch_pfn(addr);
240
241 return map_kernel_page(text_poke_addr, (pfn << PAGE_SHIFT), PAGE_KERNEL);
242}
243
244static void unmap_patch_area(unsigned long addr)
245{
246 pte_t *ptep;
247 pmd_t *pmdp;
248 pud_t *pudp;
249 p4d_t *p4dp;
250 pgd_t *pgdp;
251
252 pgdp = pgd_offset_k(addr);
253 if (WARN_ON(pgd_none(*pgdp)))
254 return;
255
256 p4dp = p4d_offset(pgdp, addr);
257 if (WARN_ON(p4d_none(*p4dp)))
258 return;
259
260 pudp = pud_offset(p4dp, addr);
261 if (WARN_ON(pud_none(*pudp)))
262 return;
263
264 pmdp = pmd_offset(pudp, addr);
265 if (WARN_ON(pmd_none(*pmdp)))
266 return;
267
268 ptep = pte_offset_kernel(pmdp, addr);
269 if (WARN_ON(pte_none(*ptep)))
270 return;
271
272 /*
273 * In hash, pte_clear flushes the tlb, in radix, we have to
274 */
275 pte_clear(&init_mm, addr, ptep);
276 flush_tlb_kernel_range(addr, addr + PAGE_SIZE);
277}
278
279static int __do_patch_instruction_mm(u32 *addr, ppc_inst_t instr)
280{
281 int err;
282 u32 *patch_addr;
283 unsigned long text_poke_addr;
284 pte_t *pte;
285 unsigned long pfn = get_patch_pfn(addr);
286 struct mm_struct *patching_mm;
287 struct mm_struct *orig_mm;
288 spinlock_t *ptl;
289
290 patching_mm = __this_cpu_read(cpu_patching_context.mm);
291 text_poke_addr = __this_cpu_read(cpu_patching_context.addr);
292 patch_addr = (u32 *)(text_poke_addr + offset_in_page(addr));
293
294 pte = get_locked_pte(patching_mm, text_poke_addr, &ptl);
295 if (!pte)
296 return -ENOMEM;
297
298 __set_pte_at(patching_mm, text_poke_addr, pte, pfn_pte(pfn, PAGE_KERNEL), 0);
299
300 /* order PTE update before use, also serves as the hwsync */
301 asm volatile("ptesync": : :"memory");
302
303 /* order context switch after arbitrary prior code */
304 isync();
305
306 orig_mm = start_using_temp_mm(patching_mm);
307
308 err = __patch_instruction(addr, instr, patch_addr);
309
310 /* context synchronisation performed by __patch_instruction (isync or exception) */
311 stop_using_temp_mm(patching_mm, orig_mm);
312
313 pte_clear(patching_mm, text_poke_addr, pte);
314 /*
315 * ptesync to order PTE update before TLB invalidation done
316 * by radix__local_flush_tlb_page_psize (in _tlbiel_va)
317 */
318 local_flush_tlb_page_psize(patching_mm, text_poke_addr, mmu_virtual_psize);
319
320 pte_unmap_unlock(pte, ptl);
321
322 return err;
323}
324
325static int __do_patch_instruction(u32 *addr, ppc_inst_t instr)
326{
327 int err;
328 u32 *patch_addr;
329 unsigned long text_poke_addr;
330 pte_t *pte;
331 unsigned long pfn = get_patch_pfn(addr);
332
333 text_poke_addr = (unsigned long)__this_cpu_read(cpu_patching_context.addr) & PAGE_MASK;
334 patch_addr = (u32 *)(text_poke_addr + offset_in_page(addr));
335
336 pte = __this_cpu_read(cpu_patching_context.pte);
337 __set_pte_at(&init_mm, text_poke_addr, pte, pfn_pte(pfn, PAGE_KERNEL), 0);
338 /* See ptesync comment in radix__set_pte_at() */
339 if (radix_enabled())
340 asm volatile("ptesync": : :"memory");
341
342 err = __patch_instruction(addr, instr, patch_addr);
343
344 pte_clear(&init_mm, text_poke_addr, pte);
345 flush_tlb_kernel_range(text_poke_addr, text_poke_addr + PAGE_SIZE);
346
347 return err;
348}
349
350int patch_instruction(u32 *addr, ppc_inst_t instr)
351{
352 int err;
353 unsigned long flags;
354
355 /*
356 * During early early boot patch_instruction is called
357 * when text_poke_area is not ready, but we still need
358 * to allow patching. We just do the plain old patching
359 */
360 if (!IS_ENABLED(CONFIG_STRICT_KERNEL_RWX) ||
361 !static_branch_likely(&poking_init_done))
362 return raw_patch_instruction(addr, instr);
363
364 local_irq_save(flags);
365 if (mm_patch_enabled())
366 err = __do_patch_instruction_mm(addr, instr);
367 else
368 err = __do_patch_instruction(addr, instr);
369 local_irq_restore(flags);
370
371 return err;
372}
373NOKPROBE_SYMBOL(patch_instruction);
374
375static int patch_memset64(u64 *addr, u64 val, size_t count)
376{
377 for (u64 *end = addr + count; addr < end; addr++)
378 __put_kernel_nofault(addr, &val, u64, failed);
379
380 return 0;
381
382failed:
383 return -EPERM;
384}
385
386static int patch_memset32(u32 *addr, u32 val, size_t count)
387{
388 for (u32 *end = addr + count; addr < end; addr++)
389 __put_kernel_nofault(addr, &val, u32, failed);
390
391 return 0;
392
393failed:
394 return -EPERM;
395}
396
397static int __patch_instructions(u32 *patch_addr, u32 *code, size_t len, bool repeat_instr)
398{
399 unsigned long start = (unsigned long)patch_addr;
400 int err;
401
402 /* Repeat instruction */
403 if (repeat_instr) {
404 ppc_inst_t instr = ppc_inst_read(code);
405
406 if (ppc_inst_prefixed(instr)) {
407 u64 val = ppc_inst_as_ulong(instr);
408
409 err = patch_memset64((u64 *)patch_addr, val, len / 8);
410 } else {
411 u32 val = ppc_inst_val(instr);
412
413 err = patch_memset32(patch_addr, val, len / 4);
414 }
415 } else {
416 err = copy_to_kernel_nofault(patch_addr, code, len);
417 }
418
419 smp_wmb(); /* smp write barrier */
420 flush_icache_range(start, start + len);
421 return err;
422}
423
424/*
425 * A page is mapped and instructions that fit the page are patched.
426 * Assumes 'len' to be (PAGE_SIZE - offset_in_page(addr)) or below.
427 */
428static int __do_patch_instructions_mm(u32 *addr, u32 *code, size_t len, bool repeat_instr)
429{
430 struct mm_struct *patching_mm, *orig_mm;
431 unsigned long pfn = get_patch_pfn(addr);
432 unsigned long text_poke_addr;
433 spinlock_t *ptl;
434 u32 *patch_addr;
435 pte_t *pte;
436 int err;
437
438 patching_mm = __this_cpu_read(cpu_patching_context.mm);
439 text_poke_addr = __this_cpu_read(cpu_patching_context.addr);
440 patch_addr = (u32 *)(text_poke_addr + offset_in_page(addr));
441
442 pte = get_locked_pte(patching_mm, text_poke_addr, &ptl);
443 if (!pte)
444 return -ENOMEM;
445
446 __set_pte_at(patching_mm, text_poke_addr, pte, pfn_pte(pfn, PAGE_KERNEL), 0);
447
448 /* order PTE update before use, also serves as the hwsync */
449 asm volatile("ptesync" ::: "memory");
450
451 /* order context switch after arbitrary prior code */
452 isync();
453
454 orig_mm = start_using_temp_mm(patching_mm);
455
456 err = __patch_instructions(patch_addr, code, len, repeat_instr);
457
458 /* context synchronisation performed by __patch_instructions */
459 stop_using_temp_mm(patching_mm, orig_mm);
460
461 pte_clear(patching_mm, text_poke_addr, pte);
462 /*
463 * ptesync to order PTE update before TLB invalidation done
464 * by radix__local_flush_tlb_page_psize (in _tlbiel_va)
465 */
466 local_flush_tlb_page_psize(patching_mm, text_poke_addr, mmu_virtual_psize);
467
468 pte_unmap_unlock(pte, ptl);
469
470 return err;
471}
472
473/*
474 * A page is mapped and instructions that fit the page are patched.
475 * Assumes 'len' to be (PAGE_SIZE - offset_in_page(addr)) or below.
476 */
477static int __do_patch_instructions(u32 *addr, u32 *code, size_t len, bool repeat_instr)
478{
479 unsigned long pfn = get_patch_pfn(addr);
480 unsigned long text_poke_addr;
481 u32 *patch_addr;
482 pte_t *pte;
483 int err;
484
485 text_poke_addr = (unsigned long)__this_cpu_read(cpu_patching_context.addr) & PAGE_MASK;
486 patch_addr = (u32 *)(text_poke_addr + offset_in_page(addr));
487
488 pte = __this_cpu_read(cpu_patching_context.pte);
489 __set_pte_at(&init_mm, text_poke_addr, pte, pfn_pte(pfn, PAGE_KERNEL), 0);
490 /* See ptesync comment in radix__set_pte_at() */
491 if (radix_enabled())
492 asm volatile("ptesync" ::: "memory");
493
494 err = __patch_instructions(patch_addr, code, len, repeat_instr);
495
496 pte_clear(&init_mm, text_poke_addr, pte);
497 flush_tlb_kernel_range(text_poke_addr, text_poke_addr + PAGE_SIZE);
498
499 return err;
500}
501
502/*
503 * Patch 'addr' with 'len' bytes of instructions from 'code'.
504 *
505 * If repeat_instr is true, the same instruction is filled for
506 * 'len' bytes.
507 */
508int patch_instructions(u32 *addr, u32 *code, size_t len, bool repeat_instr)
509{
510 while (len > 0) {
511 unsigned long flags;
512 size_t plen;
513 int err;
514
515 plen = min_t(size_t, PAGE_SIZE - offset_in_page(addr), len);
516
517 local_irq_save(flags);
518 if (mm_patch_enabled())
519 err = __do_patch_instructions_mm(addr, code, plen, repeat_instr);
520 else
521 err = __do_patch_instructions(addr, code, plen, repeat_instr);
522 local_irq_restore(flags);
523 if (err)
524 return err;
525
526 len -= plen;
527 addr = (u32 *)((unsigned long)addr + plen);
528 if (!repeat_instr)
529 code = (u32 *)((unsigned long)code + plen);
530 }
531
532 return 0;
533}
534NOKPROBE_SYMBOL(patch_instructions);
535
536int patch_branch(u32 *addr, unsigned long target, int flags)
537{
538 ppc_inst_t instr;
539
540 if (create_branch(&instr, addr, target, flags))
541 return -ERANGE;
542
543 return patch_instruction(addr, instr);
544}
545
546/*
547 * Helper to check if a given instruction is a conditional branch
548 * Derived from the conditional checks in analyse_instr()
549 */
550bool is_conditional_branch(ppc_inst_t instr)
551{
552 unsigned int opcode = ppc_inst_primary_opcode(instr);
553
554 if (opcode == 16) /* bc, bca, bcl, bcla */
555 return true;
556 if (opcode == 19) {
557 switch ((ppc_inst_val(instr) >> 1) & 0x3ff) {
558 case 16: /* bclr, bclrl */
559 case 528: /* bcctr, bcctrl */
560 case 560: /* bctar, bctarl */
561 return true;
562 }
563 }
564 return false;
565}
566NOKPROBE_SYMBOL(is_conditional_branch);
567
568int create_cond_branch(ppc_inst_t *instr, const u32 *addr,
569 unsigned long target, int flags)
570{
571 long offset;
572
573 offset = target;
574 if (! (flags & BRANCH_ABSOLUTE))
575 offset = offset - (unsigned long)addr;
576
577 /* Check we can represent the target in the instruction format */
578 if (!is_offset_in_cond_branch_range(offset))
579 return 1;
580
581 /* Mask out the flags and target, so they don't step on each other. */
582 *instr = ppc_inst(0x40000000 | (flags & 0x3FF0003) | (offset & 0xFFFC));
583
584 return 0;
585}
586
587int instr_is_relative_branch(ppc_inst_t instr)
588{
589 if (ppc_inst_val(instr) & BRANCH_ABSOLUTE)
590 return 0;
591
592 return instr_is_branch_iform(instr) || instr_is_branch_bform(instr);
593}
594
595int instr_is_relative_link_branch(ppc_inst_t instr)
596{
597 return instr_is_relative_branch(instr) && (ppc_inst_val(instr) & BRANCH_SET_LINK);
598}
599
600static unsigned long branch_iform_target(const u32 *instr)
601{
602 signed long imm;
603
604 imm = ppc_inst_val(ppc_inst_read(instr)) & 0x3FFFFFC;
605
606 /* If the top bit of the immediate value is set this is negative */
607 if (imm & 0x2000000)
608 imm -= 0x4000000;
609
610 if ((ppc_inst_val(ppc_inst_read(instr)) & BRANCH_ABSOLUTE) == 0)
611 imm += (unsigned long)instr;
612
613 return (unsigned long)imm;
614}
615
616static unsigned long branch_bform_target(const u32 *instr)
617{
618 signed long imm;
619
620 imm = ppc_inst_val(ppc_inst_read(instr)) & 0xFFFC;
621
622 /* If the top bit of the immediate value is set this is negative */
623 if (imm & 0x8000)
624 imm -= 0x10000;
625
626 if ((ppc_inst_val(ppc_inst_read(instr)) & BRANCH_ABSOLUTE) == 0)
627 imm += (unsigned long)instr;
628
629 return (unsigned long)imm;
630}
631
632unsigned long branch_target(const u32 *instr)
633{
634 if (instr_is_branch_iform(ppc_inst_read(instr)))
635 return branch_iform_target(instr);
636 else if (instr_is_branch_bform(ppc_inst_read(instr)))
637 return branch_bform_target(instr);
638
639 return 0;
640}
641
642int translate_branch(ppc_inst_t *instr, const u32 *dest, const u32 *src)
643{
644 unsigned long target;
645 target = branch_target(src);
646
647 if (instr_is_branch_iform(ppc_inst_read(src)))
648 return create_branch(instr, dest, target,
649 ppc_inst_val(ppc_inst_read(src)));
650 else if (instr_is_branch_bform(ppc_inst_read(src)))
651 return create_cond_branch(instr, dest, target,
652 ppc_inst_val(ppc_inst_read(src)));
653
654 return 1;
655}