Merge branch 'topic/objtool' into next
[linux-block.git] / arch / powerpc / lib / code-patching.c
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
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
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
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  */
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
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
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
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
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
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         this_cpu_write(cpu_patching_context.pte, pte);
182
183         return 0;
184
185 fail_no_pte:
186         put_patching_mm(mm, addr);
187 fail_no_mm:
188         return -ENOMEM;
189 }
190
191 static 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         this_cpu_write(cpu_patching_context.pte, NULL);
199
200         return 0;
201 }
202
203 static __ro_after_init DEFINE_STATIC_KEY_FALSE(poking_init_done);
204
205 void __init poking_init(void)
206 {
207         int ret;
208
209         if (!IS_ENABLED(CONFIG_STRICT_KERNEL_RWX))
210                 return;
211
212         if (mm_patch_enabled())
213                 ret = cpuhp_setup_state(CPUHP_AP_ONLINE_DYN,
214                                         "powerpc/text_poke_mm:online",
215                                         text_area_cpu_up_mm,
216                                         text_area_cpu_down_mm);
217         else
218                 ret = cpuhp_setup_state(CPUHP_AP_ONLINE_DYN,
219                                         "powerpc/text_poke:online",
220                                         text_area_cpu_up,
221                                         text_area_cpu_down);
222
223         /* cpuhp_setup_state returns >= 0 on success */
224         if (WARN_ON(ret < 0))
225                 return;
226
227         static_branch_enable(&poking_init_done);
228 }
229
230 static unsigned long get_patch_pfn(void *addr)
231 {
232         if (IS_ENABLED(CONFIG_MODULES) && is_vmalloc_or_module_addr(addr))
233                 return vmalloc_to_pfn(addr);
234         else
235                 return __pa_symbol(addr) >> PAGE_SHIFT;
236 }
237
238 /*
239  * This can be called for kernel text or a module.
240  */
241 static int map_patch_area(void *addr, unsigned long text_poke_addr)
242 {
243         unsigned long pfn = get_patch_pfn(addr);
244
245         return map_kernel_page(text_poke_addr, (pfn << PAGE_SHIFT), PAGE_KERNEL);
246 }
247
248 static void unmap_patch_area(unsigned long addr)
249 {
250         pte_t *ptep;
251         pmd_t *pmdp;
252         pud_t *pudp;
253         p4d_t *p4dp;
254         pgd_t *pgdp;
255
256         pgdp = pgd_offset_k(addr);
257         if (WARN_ON(pgd_none(*pgdp)))
258                 return;
259
260         p4dp = p4d_offset(pgdp, addr);
261         if (WARN_ON(p4d_none(*p4dp)))
262                 return;
263
264         pudp = pud_offset(p4dp, addr);
265         if (WARN_ON(pud_none(*pudp)))
266                 return;
267
268         pmdp = pmd_offset(pudp, addr);
269         if (WARN_ON(pmd_none(*pmdp)))
270                 return;
271
272         ptep = pte_offset_kernel(pmdp, addr);
273         if (WARN_ON(pte_none(*ptep)))
274                 return;
275
276         /*
277          * In hash, pte_clear flushes the tlb, in radix, we have to
278          */
279         pte_clear(&init_mm, addr, ptep);
280         flush_tlb_kernel_range(addr, addr + PAGE_SIZE);
281 }
282
283 static int __do_patch_instruction_mm(u32 *addr, ppc_inst_t instr)
284 {
285         int err;
286         u32 *patch_addr;
287         unsigned long text_poke_addr;
288         pte_t *pte;
289         unsigned long pfn = get_patch_pfn(addr);
290         struct mm_struct *patching_mm;
291         struct mm_struct *orig_mm;
292
293         patching_mm = __this_cpu_read(cpu_patching_context.mm);
294         pte = __this_cpu_read(cpu_patching_context.pte);
295         text_poke_addr = __this_cpu_read(cpu_patching_context.addr);
296         patch_addr = (u32 *)(text_poke_addr + offset_in_page(addr));
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         /* hwsync performed by __patch_instruction (sync) if successful */
311         if (err)
312                 mb();  /* sync */
313
314         /* context synchronisation performed by __patch_instruction (isync or exception) */
315         stop_using_temp_mm(patching_mm, orig_mm);
316
317         pte_clear(patching_mm, text_poke_addr, pte);
318         /*
319          * ptesync to order PTE update before TLB invalidation done
320          * by radix__local_flush_tlb_page_psize (in _tlbiel_va)
321          */
322         local_flush_tlb_page_psize(patching_mm, text_poke_addr, mmu_virtual_psize);
323
324         return err;
325 }
326
327 static int __do_patch_instruction(u32 *addr, ppc_inst_t instr)
328 {
329         int err;
330         u32 *patch_addr;
331         unsigned long text_poke_addr;
332         pte_t *pte;
333         unsigned long pfn = get_patch_pfn(addr);
334
335         text_poke_addr = (unsigned long)__this_cpu_read(cpu_patching_context.addr) & PAGE_MASK;
336         patch_addr = (u32 *)(text_poke_addr + offset_in_page(addr));
337
338         pte = __this_cpu_read(cpu_patching_context.pte);
339         __set_pte_at(&init_mm, text_poke_addr, pte, pfn_pte(pfn, PAGE_KERNEL), 0);
340         /* See ptesync comment in radix__set_pte_at() */
341         if (radix_enabled())
342                 asm volatile("ptesync": : :"memory");
343
344         err = __patch_instruction(addr, instr, patch_addr);
345
346         pte_clear(&init_mm, text_poke_addr, pte);
347         flush_tlb_kernel_range(text_poke_addr, text_poke_addr + PAGE_SIZE);
348
349         return err;
350 }
351
352 int patch_instruction(u32 *addr, ppc_inst_t instr)
353 {
354         int err;
355         unsigned long flags;
356
357         /*
358          * During early early boot patch_instruction is called
359          * when text_poke_area is not ready, but we still need
360          * to allow patching. We just do the plain old patching
361          */
362         if (!IS_ENABLED(CONFIG_STRICT_KERNEL_RWX) ||
363             !static_branch_likely(&poking_init_done))
364                 return raw_patch_instruction(addr, instr);
365
366         local_irq_save(flags);
367         if (mm_patch_enabled())
368                 err = __do_patch_instruction_mm(addr, instr);
369         else
370                 err = __do_patch_instruction(addr, instr);
371         local_irq_restore(flags);
372
373         return err;
374 }
375 NOKPROBE_SYMBOL(patch_instruction);
376
377 int patch_branch(u32 *addr, unsigned long target, int flags)
378 {
379         ppc_inst_t instr;
380
381         if (create_branch(&instr, addr, target, flags))
382                 return -ERANGE;
383
384         return patch_instruction(addr, instr);
385 }
386
387 /*
388  * Helper to check if a given instruction is a conditional branch
389  * Derived from the conditional checks in analyse_instr()
390  */
391 bool is_conditional_branch(ppc_inst_t instr)
392 {
393         unsigned int opcode = ppc_inst_primary_opcode(instr);
394
395         if (opcode == 16)       /* bc, bca, bcl, bcla */
396                 return true;
397         if (opcode == 19) {
398                 switch ((ppc_inst_val(instr) >> 1) & 0x3ff) {
399                 case 16:        /* bclr, bclrl */
400                 case 528:       /* bcctr, bcctrl */
401                 case 560:       /* bctar, bctarl */
402                         return true;
403                 }
404         }
405         return false;
406 }
407 NOKPROBE_SYMBOL(is_conditional_branch);
408
409 int create_cond_branch(ppc_inst_t *instr, const u32 *addr,
410                        unsigned long target, int flags)
411 {
412         long offset;
413
414         offset = target;
415         if (! (flags & BRANCH_ABSOLUTE))
416                 offset = offset - (unsigned long)addr;
417
418         /* Check we can represent the target in the instruction format */
419         if (!is_offset_in_cond_branch_range(offset))
420                 return 1;
421
422         /* Mask out the flags and target, so they don't step on each other. */
423         *instr = ppc_inst(0x40000000 | (flags & 0x3FF0003) | (offset & 0xFFFC));
424
425         return 0;
426 }
427
428 int instr_is_relative_branch(ppc_inst_t instr)
429 {
430         if (ppc_inst_val(instr) & BRANCH_ABSOLUTE)
431                 return 0;
432
433         return instr_is_branch_iform(instr) || instr_is_branch_bform(instr);
434 }
435
436 int instr_is_relative_link_branch(ppc_inst_t instr)
437 {
438         return instr_is_relative_branch(instr) && (ppc_inst_val(instr) & BRANCH_SET_LINK);
439 }
440
441 static unsigned long branch_iform_target(const u32 *instr)
442 {
443         signed long imm;
444
445         imm = ppc_inst_val(ppc_inst_read(instr)) & 0x3FFFFFC;
446
447         /* If the top bit of the immediate value is set this is negative */
448         if (imm & 0x2000000)
449                 imm -= 0x4000000;
450
451         if ((ppc_inst_val(ppc_inst_read(instr)) & BRANCH_ABSOLUTE) == 0)
452                 imm += (unsigned long)instr;
453
454         return (unsigned long)imm;
455 }
456
457 static unsigned long branch_bform_target(const u32 *instr)
458 {
459         signed long imm;
460
461         imm = ppc_inst_val(ppc_inst_read(instr)) & 0xFFFC;
462
463         /* If the top bit of the immediate value is set this is negative */
464         if (imm & 0x8000)
465                 imm -= 0x10000;
466
467         if ((ppc_inst_val(ppc_inst_read(instr)) & BRANCH_ABSOLUTE) == 0)
468                 imm += (unsigned long)instr;
469
470         return (unsigned long)imm;
471 }
472
473 unsigned long branch_target(const u32 *instr)
474 {
475         if (instr_is_branch_iform(ppc_inst_read(instr)))
476                 return branch_iform_target(instr);
477         else if (instr_is_branch_bform(ppc_inst_read(instr)))
478                 return branch_bform_target(instr);
479
480         return 0;
481 }
482
483 int translate_branch(ppc_inst_t *instr, const u32 *dest, const u32 *src)
484 {
485         unsigned long target;
486         target = branch_target(src);
487
488         if (instr_is_branch_iform(ppc_inst_read(src)))
489                 return create_branch(instr, dest, target,
490                                      ppc_inst_val(ppc_inst_read(src)));
491         else if (instr_is_branch_bform(ppc_inst_read(src)))
492                 return create_cond_branch(instr, dest, target,
493                                           ppc_inst_val(ppc_inst_read(src)));
494
495         return 1;
496 }