Loading...
1// SPDX-License-Identifier: GPL-2.0-or-later
2/* Kernel module help for PPC64.
3 Copyright (C) 2001, 2003 Rusty Russell IBM Corporation.
4
5*/
6
7#define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
8
9#include <linux/module.h>
10#include <linux/elf.h>
11#include <linux/moduleloader.h>
12#include <linux/err.h>
13#include <linux/vmalloc.h>
14#include <linux/ftrace.h>
15#include <linux/bug.h>
16#include <linux/uaccess.h>
17#include <asm/module.h>
18#include <asm/firmware.h>
19#include <asm/code-patching.h>
20#include <linux/sort.h>
21#include <asm/setup.h>
22#include <asm/sections.h>
23#include <asm/inst.h>
24
25/* FIXME: We don't do .init separately. To do this, we'd need to have
26 a separate r2 value in the init and core section, and stub between
27 them, too.
28
29 Using a magic allocator which places modules within 32MB solves
30 this, and makes other things simpler. Anton?
31 --RR. */
32
33#ifdef PPC64_ELF_ABI_v2
34
35/* An address is simply the address of the function. */
36typedef unsigned long func_desc_t;
37
38static func_desc_t func_desc(unsigned long addr)
39{
40 return addr;
41}
42static unsigned long func_addr(unsigned long addr)
43{
44 return addr;
45}
46static unsigned long stub_func_addr(func_desc_t func)
47{
48 return func;
49}
50
51/* PowerPC64 specific values for the Elf64_Sym st_other field. */
52#define STO_PPC64_LOCAL_BIT 5
53#define STO_PPC64_LOCAL_MASK (7 << STO_PPC64_LOCAL_BIT)
54#define PPC64_LOCAL_ENTRY_OFFSET(other) \
55 (((1 << (((other) & STO_PPC64_LOCAL_MASK) >> STO_PPC64_LOCAL_BIT)) >> 2) << 2)
56
57static unsigned int local_entry_offset(const Elf64_Sym *sym)
58{
59 /* sym->st_other indicates offset to local entry point
60 * (otherwise it will assume r12 is the address of the start
61 * of function and try to derive r2 from it). */
62 return PPC64_LOCAL_ENTRY_OFFSET(sym->st_other);
63}
64#else
65
66/* An address is address of the OPD entry, which contains address of fn. */
67typedef struct ppc64_opd_entry func_desc_t;
68
69static func_desc_t func_desc(unsigned long addr)
70{
71 return *(struct ppc64_opd_entry *)addr;
72}
73static unsigned long func_addr(unsigned long addr)
74{
75 return func_desc(addr).funcaddr;
76}
77static unsigned long stub_func_addr(func_desc_t func)
78{
79 return func.funcaddr;
80}
81static unsigned int local_entry_offset(const Elf64_Sym *sym)
82{
83 return 0;
84}
85
86void *dereference_module_function_descriptor(struct module *mod, void *ptr)
87{
88 if (ptr < (void *)mod->arch.start_opd ||
89 ptr >= (void *)mod->arch.end_opd)
90 return ptr;
91
92 return dereference_function_descriptor(ptr);
93}
94#endif
95
96#define STUB_MAGIC 0x73747562 /* stub */
97
98/* Like PPC32, we need little trampolines to do > 24-bit jumps (into
99 the kernel itself). But on PPC64, these need to be used for every
100 jump, actually, to reset r2 (TOC+0x8000). */
101struct ppc64_stub_entry
102{
103 /* 28 byte jump instruction sequence (7 instructions). We only
104 * need 6 instructions on ABIv2 but we always allocate 7 so
105 * so we don't have to modify the trampoline load instruction. */
106 u32 jump[7];
107 /* Used by ftrace to identify stubs */
108 u32 magic;
109 /* Data for the above code */
110 func_desc_t funcdata;
111};
112
113/*
114 * PPC64 uses 24 bit jumps, but we need to jump into other modules or
115 * the kernel which may be further. So we jump to a stub.
116 *
117 * For ELFv1 we need to use this to set up the new r2 value (aka TOC
118 * pointer). For ELFv2 it's the callee's responsibility to set up the
119 * new r2, but for both we need to save the old r2.
120 *
121 * We could simply patch the new r2 value and function pointer into
122 * the stub, but it's significantly shorter to put these values at the
123 * end of the stub code, and patch the stub address (32-bits relative
124 * to the TOC ptr, r2) into the stub.
125 *
126 * addis r11,r2, <high>
127 * addi r11,r11, <low>
128 * std r2,R2_STACK_OFFSET(r1)
129 * ld r12,32(r11)
130 * ld r2,40(r11)
131 * mtctr r12
132 * bctr
133 */
134static u32 ppc64_stub_insns[] = {
135 PPC_INST_ADDIS | __PPC_RT(R11) | __PPC_RA(R2),
136 PPC_INST_ADDI | __PPC_RT(R11) | __PPC_RA(R11),
137 /* Save current r2 value in magic place on the stack. */
138 PPC_INST_STD | __PPC_RS(R2) | __PPC_RA(R1) | R2_STACK_OFFSET,
139 PPC_INST_LD | __PPC_RT(R12) | __PPC_RA(R11) | 32,
140#ifdef PPC64_ELF_ABI_v1
141 /* Set up new r2 from function descriptor */
142 PPC_INST_LD | __PPC_RT(R2) | __PPC_RA(R11) | 40,
143#endif
144 PPC_INST_MTCTR | __PPC_RS(R12),
145 PPC_INST_BCTR,
146};
147
148/* Count how many different 24-bit relocations (different symbol,
149 different addend) */
150static unsigned int count_relocs(const Elf64_Rela *rela, unsigned int num)
151{
152 unsigned int i, r_info, r_addend, _count_relocs;
153
154 /* FIXME: Only count external ones --RR */
155 _count_relocs = 0;
156 r_info = 0;
157 r_addend = 0;
158 for (i = 0; i < num; i++)
159 /* Only count 24-bit relocs, others don't need stubs */
160 if (ELF64_R_TYPE(rela[i].r_info) == R_PPC_REL24 &&
161 (r_info != ELF64_R_SYM(rela[i].r_info) ||
162 r_addend != rela[i].r_addend)) {
163 _count_relocs++;
164 r_info = ELF64_R_SYM(rela[i].r_info);
165 r_addend = rela[i].r_addend;
166 }
167
168 return _count_relocs;
169}
170
171static int relacmp(const void *_x, const void *_y)
172{
173 const Elf64_Rela *x, *y;
174
175 y = (Elf64_Rela *)_x;
176 x = (Elf64_Rela *)_y;
177
178 /* Compare the entire r_info (as opposed to ELF64_R_SYM(r_info) only) to
179 * make the comparison cheaper/faster. It won't affect the sorting or
180 * the counting algorithms' performance
181 */
182 if (x->r_info < y->r_info)
183 return -1;
184 else if (x->r_info > y->r_info)
185 return 1;
186 else if (x->r_addend < y->r_addend)
187 return -1;
188 else if (x->r_addend > y->r_addend)
189 return 1;
190 else
191 return 0;
192}
193
194/* Get size of potential trampolines required. */
195static unsigned long get_stubs_size(const Elf64_Ehdr *hdr,
196 const Elf64_Shdr *sechdrs)
197{
198 /* One extra reloc so it's always 0-funcaddr terminated */
199 unsigned long relocs = 1;
200 unsigned i;
201
202 /* Every relocated section... */
203 for (i = 1; i < hdr->e_shnum; i++) {
204 if (sechdrs[i].sh_type == SHT_RELA) {
205 pr_debug("Found relocations in section %u\n", i);
206 pr_debug("Ptr: %p. Number: %Lu\n",
207 (void *)sechdrs[i].sh_addr,
208 sechdrs[i].sh_size / sizeof(Elf64_Rela));
209
210 /* Sort the relocation information based on a symbol and
211 * addend key. This is a stable O(n*log n) complexity
212 * alogrithm but it will reduce the complexity of
213 * count_relocs() to linear complexity O(n)
214 */
215 sort((void *)sechdrs[i].sh_addr,
216 sechdrs[i].sh_size / sizeof(Elf64_Rela),
217 sizeof(Elf64_Rela), relacmp, NULL);
218
219 relocs += count_relocs((void *)sechdrs[i].sh_addr,
220 sechdrs[i].sh_size
221 / sizeof(Elf64_Rela));
222 }
223 }
224
225#ifdef CONFIG_DYNAMIC_FTRACE
226 /* make the trampoline to the ftrace_caller */
227 relocs++;
228#ifdef CONFIG_DYNAMIC_FTRACE_WITH_REGS
229 /* an additional one for ftrace_regs_caller */
230 relocs++;
231#endif
232#endif
233
234 pr_debug("Looks like a total of %lu stubs, max\n", relocs);
235 return relocs * sizeof(struct ppc64_stub_entry);
236}
237
238/* Still needed for ELFv2, for .TOC. */
239static void dedotify_versions(struct modversion_info *vers,
240 unsigned long size)
241{
242 struct modversion_info *end;
243
244 for (end = (void *)vers + size; vers < end; vers++)
245 if (vers->name[0] == '.') {
246 memmove(vers->name, vers->name+1, strlen(vers->name));
247 }
248}
249
250/*
251 * Undefined symbols which refer to .funcname, hack to funcname. Make .TOC.
252 * seem to be defined (value set later).
253 */
254static void dedotify(Elf64_Sym *syms, unsigned int numsyms, char *strtab)
255{
256 unsigned int i;
257
258 for (i = 1; i < numsyms; i++) {
259 if (syms[i].st_shndx == SHN_UNDEF) {
260 char *name = strtab + syms[i].st_name;
261 if (name[0] == '.') {
262 if (strcmp(name+1, "TOC.") == 0)
263 syms[i].st_shndx = SHN_ABS;
264 syms[i].st_name++;
265 }
266 }
267 }
268}
269
270static Elf64_Sym *find_dot_toc(Elf64_Shdr *sechdrs,
271 const char *strtab,
272 unsigned int symindex)
273{
274 unsigned int i, numsyms;
275 Elf64_Sym *syms;
276
277 syms = (Elf64_Sym *)sechdrs[symindex].sh_addr;
278 numsyms = sechdrs[symindex].sh_size / sizeof(Elf64_Sym);
279
280 for (i = 1; i < numsyms; i++) {
281 if (syms[i].st_shndx == SHN_ABS
282 && strcmp(strtab + syms[i].st_name, "TOC.") == 0)
283 return &syms[i];
284 }
285 return NULL;
286}
287
288int module_frob_arch_sections(Elf64_Ehdr *hdr,
289 Elf64_Shdr *sechdrs,
290 char *secstrings,
291 struct module *me)
292{
293 unsigned int i;
294
295 /* Find .toc and .stubs sections, symtab and strtab */
296 for (i = 1; i < hdr->e_shnum; i++) {
297 char *p;
298 if (strcmp(secstrings + sechdrs[i].sh_name, ".stubs") == 0)
299 me->arch.stubs_section = i;
300 else if (strcmp(secstrings + sechdrs[i].sh_name, ".toc") == 0) {
301 me->arch.toc_section = i;
302 if (sechdrs[i].sh_addralign < 8)
303 sechdrs[i].sh_addralign = 8;
304 }
305 else if (strcmp(secstrings+sechdrs[i].sh_name,"__versions")==0)
306 dedotify_versions((void *)hdr + sechdrs[i].sh_offset,
307 sechdrs[i].sh_size);
308
309 /* We don't handle .init for the moment: rename to _init */
310 while ((p = strstr(secstrings + sechdrs[i].sh_name, ".init")))
311 p[0] = '_';
312
313 if (sechdrs[i].sh_type == SHT_SYMTAB)
314 dedotify((void *)hdr + sechdrs[i].sh_offset,
315 sechdrs[i].sh_size / sizeof(Elf64_Sym),
316 (void *)hdr
317 + sechdrs[sechdrs[i].sh_link].sh_offset);
318 }
319
320 if (!me->arch.stubs_section) {
321 pr_err("%s: doesn't contain .stubs.\n", me->name);
322 return -ENOEXEC;
323 }
324
325 /* If we don't have a .toc, just use .stubs. We need to set r2
326 to some reasonable value in case the module calls out to
327 other functions via a stub, or if a function pointer escapes
328 the module by some means. */
329 if (!me->arch.toc_section)
330 me->arch.toc_section = me->arch.stubs_section;
331
332 /* Override the stubs size */
333 sechdrs[me->arch.stubs_section].sh_size = get_stubs_size(hdr, sechdrs);
334 return 0;
335}
336
337#ifdef CONFIG_MPROFILE_KERNEL
338
339#define PACATOC offsetof(struct paca_struct, kernel_toc)
340
341/*
342 * ld r12,PACATOC(r13)
343 * addis r12,r12,<high>
344 * addi r12,r12,<low>
345 * mtctr r12
346 * bctr
347 */
348static u32 stub_insns[] = {
349 PPC_INST_LD | __PPC_RT(R12) | __PPC_RA(R13) | PACATOC,
350 PPC_INST_ADDIS | __PPC_RT(R12) | __PPC_RA(R12),
351 PPC_INST_ADDI | __PPC_RT(R12) | __PPC_RA(R12),
352 PPC_INST_MTCTR | __PPC_RS(R12),
353 PPC_INST_BCTR,
354};
355
356/*
357 * For mprofile-kernel we use a special stub for ftrace_caller() because we
358 * can't rely on r2 containing this module's TOC when we enter the stub.
359 *
360 * That can happen if the function calling us didn't need to use the toc. In
361 * that case it won't have setup r2, and the r2 value will be either the
362 * kernel's toc, or possibly another modules toc.
363 *
364 * To deal with that this stub uses the kernel toc, which is always accessible
365 * via the paca (in r13). The target (ftrace_caller()) is responsible for
366 * saving and restoring the toc before returning.
367 */
368static inline int create_ftrace_stub(struct ppc64_stub_entry *entry,
369 unsigned long addr,
370 struct module *me)
371{
372 long reladdr;
373
374 memcpy(entry->jump, stub_insns, sizeof(stub_insns));
375
376 /* Stub uses address relative to kernel toc (from the paca) */
377 reladdr = addr - kernel_toc_addr();
378 if (reladdr > 0x7FFFFFFF || reladdr < -(0x80000000L)) {
379 pr_err("%s: Address of %ps out of range of kernel_toc.\n",
380 me->name, (void *)addr);
381 return 0;
382 }
383
384 entry->jump[1] |= PPC_HA(reladdr);
385 entry->jump[2] |= PPC_LO(reladdr);
386
387 /* Eventhough we don't use funcdata in the stub, it's needed elsewhere. */
388 entry->funcdata = func_desc(addr);
389 entry->magic = STUB_MAGIC;
390
391 return 1;
392}
393
394static bool is_mprofile_ftrace_call(const char *name)
395{
396 if (!strcmp("_mcount", name))
397 return true;
398#ifdef CONFIG_DYNAMIC_FTRACE
399 if (!strcmp("ftrace_caller", name))
400 return true;
401#ifdef CONFIG_DYNAMIC_FTRACE_WITH_REGS
402 if (!strcmp("ftrace_regs_caller", name))
403 return true;
404#endif
405#endif
406
407 return false;
408}
409#else
410static inline int create_ftrace_stub(struct ppc64_stub_entry *entry,
411 unsigned long addr,
412 struct module *me)
413{
414 return 0;
415}
416
417static bool is_mprofile_ftrace_call(const char *name)
418{
419 return false;
420}
421#endif
422
423/*
424 * r2 is the TOC pointer: it actually points 0x8000 into the TOC (this gives the
425 * value maximum span in an instruction which uses a signed offset). Round down
426 * to a 256 byte boundary for the odd case where we are setting up r2 without a
427 * .toc section.
428 */
429static inline unsigned long my_r2(const Elf64_Shdr *sechdrs, struct module *me)
430{
431 return (sechdrs[me->arch.toc_section].sh_addr & ~0xfful) + 0x8000;
432}
433
434/* Patch stub to reference function and correct r2 value. */
435static inline int create_stub(const Elf64_Shdr *sechdrs,
436 struct ppc64_stub_entry *entry,
437 unsigned long addr,
438 struct module *me,
439 const char *name)
440{
441 long reladdr;
442
443 if (is_mprofile_ftrace_call(name))
444 return create_ftrace_stub(entry, addr, me);
445
446 memcpy(entry->jump, ppc64_stub_insns, sizeof(ppc64_stub_insns));
447
448 /* Stub uses address relative to r2. */
449 reladdr = (unsigned long)entry - my_r2(sechdrs, me);
450 if (reladdr > 0x7FFFFFFF || reladdr < -(0x80000000L)) {
451 pr_err("%s: Address %p of stub out of range of %p.\n",
452 me->name, (void *)reladdr, (void *)my_r2);
453 return 0;
454 }
455 pr_debug("Stub %p get data from reladdr %li\n", entry, reladdr);
456
457 entry->jump[0] |= PPC_HA(reladdr);
458 entry->jump[1] |= PPC_LO(reladdr);
459 entry->funcdata = func_desc(addr);
460 entry->magic = STUB_MAGIC;
461
462 return 1;
463}
464
465/* Create stub to jump to function described in this OPD/ptr: we need the
466 stub to set up the TOC ptr (r2) for the function. */
467static unsigned long stub_for_addr(const Elf64_Shdr *sechdrs,
468 unsigned long addr,
469 struct module *me,
470 const char *name)
471{
472 struct ppc64_stub_entry *stubs;
473 unsigned int i, num_stubs;
474
475 num_stubs = sechdrs[me->arch.stubs_section].sh_size / sizeof(*stubs);
476
477 /* Find this stub, or if that fails, the next avail. entry */
478 stubs = (void *)sechdrs[me->arch.stubs_section].sh_addr;
479 for (i = 0; stub_func_addr(stubs[i].funcdata); i++) {
480 if (WARN_ON(i >= num_stubs))
481 return 0;
482
483 if (stub_func_addr(stubs[i].funcdata) == func_addr(addr))
484 return (unsigned long)&stubs[i];
485 }
486
487 if (!create_stub(sechdrs, &stubs[i], addr, me, name))
488 return 0;
489
490 return (unsigned long)&stubs[i];
491}
492
493/* We expect a noop next: if it is, replace it with instruction to
494 restore r2. */
495static int restore_r2(const char *name, u32 *instruction, struct module *me)
496{
497 u32 *prev_insn = instruction - 1;
498
499 if (is_mprofile_ftrace_call(name))
500 return 1;
501
502 /*
503 * Make sure the branch isn't a sibling call. Sibling calls aren't
504 * "link" branches and they don't return, so they don't need the r2
505 * restore afterwards.
506 */
507 if (!instr_is_relative_link_branch(ppc_inst(*prev_insn)))
508 return 1;
509
510 if (*instruction != PPC_INST_NOP) {
511 pr_err("%s: Expected nop after call, got %08x at %pS\n",
512 me->name, *instruction, instruction);
513 return 0;
514 }
515 /* ld r2,R2_STACK_OFFSET(r1) */
516 *instruction = PPC_INST_LD_TOC;
517 return 1;
518}
519
520int apply_relocate_add(Elf64_Shdr *sechdrs,
521 const char *strtab,
522 unsigned int symindex,
523 unsigned int relsec,
524 struct module *me)
525{
526 unsigned int i;
527 Elf64_Rela *rela = (void *)sechdrs[relsec].sh_addr;
528 Elf64_Sym *sym;
529 unsigned long *location;
530 unsigned long value;
531
532 pr_debug("Applying ADD relocate section %u to %u\n", relsec,
533 sechdrs[relsec].sh_info);
534
535 /* First time we're called, we can fix up .TOC. */
536 if (!me->arch.toc_fixed) {
537 sym = find_dot_toc(sechdrs, strtab, symindex);
538 /* It's theoretically possible that a module doesn't want a
539 * .TOC. so don't fail it just for that. */
540 if (sym)
541 sym->st_value = my_r2(sechdrs, me);
542 me->arch.toc_fixed = true;
543 }
544
545 for (i = 0; i < sechdrs[relsec].sh_size / sizeof(*rela); i++) {
546 /* This is where to make the change */
547 location = (void *)sechdrs[sechdrs[relsec].sh_info].sh_addr
548 + rela[i].r_offset;
549 /* This is the symbol it is referring to */
550 sym = (Elf64_Sym *)sechdrs[symindex].sh_addr
551 + ELF64_R_SYM(rela[i].r_info);
552
553 pr_debug("RELOC at %p: %li-type as %s (0x%lx) + %li\n",
554 location, (long)ELF64_R_TYPE(rela[i].r_info),
555 strtab + sym->st_name, (unsigned long)sym->st_value,
556 (long)rela[i].r_addend);
557
558 /* `Everything is relative'. */
559 value = sym->st_value + rela[i].r_addend;
560
561 switch (ELF64_R_TYPE(rela[i].r_info)) {
562 case R_PPC64_ADDR32:
563 /* Simply set it */
564 *(u32 *)location = value;
565 break;
566
567 case R_PPC64_ADDR64:
568 /* Simply set it */
569 *(unsigned long *)location = value;
570 break;
571
572 case R_PPC64_TOC:
573 *(unsigned long *)location = my_r2(sechdrs, me);
574 break;
575
576 case R_PPC64_TOC16:
577 /* Subtract TOC pointer */
578 value -= my_r2(sechdrs, me);
579 if (value + 0x8000 > 0xffff) {
580 pr_err("%s: bad TOC16 relocation (0x%lx)\n",
581 me->name, value);
582 return -ENOEXEC;
583 }
584 *((uint16_t *) location)
585 = (*((uint16_t *) location) & ~0xffff)
586 | (value & 0xffff);
587 break;
588
589 case R_PPC64_TOC16_LO:
590 /* Subtract TOC pointer */
591 value -= my_r2(sechdrs, me);
592 *((uint16_t *) location)
593 = (*((uint16_t *) location) & ~0xffff)
594 | (value & 0xffff);
595 break;
596
597 case R_PPC64_TOC16_DS:
598 /* Subtract TOC pointer */
599 value -= my_r2(sechdrs, me);
600 if ((value & 3) != 0 || value + 0x8000 > 0xffff) {
601 pr_err("%s: bad TOC16_DS relocation (0x%lx)\n",
602 me->name, value);
603 return -ENOEXEC;
604 }
605 *((uint16_t *) location)
606 = (*((uint16_t *) location) & ~0xfffc)
607 | (value & 0xfffc);
608 break;
609
610 case R_PPC64_TOC16_LO_DS:
611 /* Subtract TOC pointer */
612 value -= my_r2(sechdrs, me);
613 if ((value & 3) != 0) {
614 pr_err("%s: bad TOC16_LO_DS relocation (0x%lx)\n",
615 me->name, value);
616 return -ENOEXEC;
617 }
618 *((uint16_t *) location)
619 = (*((uint16_t *) location) & ~0xfffc)
620 | (value & 0xfffc);
621 break;
622
623 case R_PPC64_TOC16_HA:
624 /* Subtract TOC pointer */
625 value -= my_r2(sechdrs, me);
626 value = ((value + 0x8000) >> 16);
627 *((uint16_t *) location)
628 = (*((uint16_t *) location) & ~0xffff)
629 | (value & 0xffff);
630 break;
631
632 case R_PPC_REL24:
633 /* FIXME: Handle weak symbols here --RR */
634 if (sym->st_shndx == SHN_UNDEF ||
635 sym->st_shndx == SHN_LIVEPATCH) {
636 /* External: go via stub */
637 value = stub_for_addr(sechdrs, value, me,
638 strtab + sym->st_name);
639 if (!value)
640 return -ENOENT;
641 if (!restore_r2(strtab + sym->st_name,
642 (u32 *)location + 1, me))
643 return -ENOEXEC;
644 } else
645 value += local_entry_offset(sym);
646
647 /* Convert value to relative */
648 value -= (unsigned long)location;
649 if (value + 0x2000000 > 0x3ffffff || (value & 3) != 0){
650 pr_err("%s: REL24 %li out of range!\n",
651 me->name, (long int)value);
652 return -ENOEXEC;
653 }
654
655 /* Only replace bits 2 through 26 */
656 *(uint32_t *)location
657 = (*(uint32_t *)location & ~0x03fffffc)
658 | (value & 0x03fffffc);
659 break;
660
661 case R_PPC64_REL64:
662 /* 64 bits relative (used by features fixups) */
663 *location = value - (unsigned long)location;
664 break;
665
666 case R_PPC64_REL32:
667 /* 32 bits relative (used by relative exception tables) */
668 /* Convert value to relative */
669 value -= (unsigned long)location;
670 if (value + 0x80000000 > 0xffffffff) {
671 pr_err("%s: REL32 %li out of range!\n",
672 me->name, (long int)value);
673 return -ENOEXEC;
674 }
675 *(u32 *)location = value;
676 break;
677
678 case R_PPC64_TOCSAVE:
679 /*
680 * Marker reloc indicates we don't have to save r2.
681 * That would only save us one instruction, so ignore
682 * it.
683 */
684 break;
685
686 case R_PPC64_ENTRY:
687 /*
688 * Optimize ELFv2 large code model entry point if
689 * the TOC is within 2GB range of current location.
690 */
691 value = my_r2(sechdrs, me) - (unsigned long)location;
692 if (value + 0x80008000 > 0xffffffff)
693 break;
694 /*
695 * Check for the large code model prolog sequence:
696 * ld r2, ...(r12)
697 * add r2, r2, r12
698 */
699 if ((((uint32_t *)location)[0] & ~0xfffc) !=
700 (PPC_INST_LD | __PPC_RT(R2) | __PPC_RA(R12)))
701 break;
702 if (((uint32_t *)location)[1] !=
703 (PPC_INST_ADD | __PPC_RT(R2) | __PPC_RA(R2) | __PPC_RB(R12)))
704 break;
705 /*
706 * If found, replace it with:
707 * addis r2, r12, (.TOC.-func)@ha
708 * addi r2, r2, (.TOC.-func)@l
709 */
710 ((uint32_t *)location)[0] = PPC_INST_ADDIS | __PPC_RT(R2) |
711 __PPC_RA(R12) | PPC_HA(value);
712 ((uint32_t *)location)[1] = PPC_INST_ADDI | __PPC_RT(R2) |
713 __PPC_RA(R2) | PPC_LO(value);
714 break;
715
716 case R_PPC64_REL16_HA:
717 /* Subtract location pointer */
718 value -= (unsigned long)location;
719 value = ((value + 0x8000) >> 16);
720 *((uint16_t *) location)
721 = (*((uint16_t *) location) & ~0xffff)
722 | (value & 0xffff);
723 break;
724
725 case R_PPC64_REL16_LO:
726 /* Subtract location pointer */
727 value -= (unsigned long)location;
728 *((uint16_t *) location)
729 = (*((uint16_t *) location) & ~0xffff)
730 | (value & 0xffff);
731 break;
732
733 default:
734 pr_err("%s: Unknown ADD relocation: %lu\n",
735 me->name,
736 (unsigned long)ELF64_R_TYPE(rela[i].r_info));
737 return -ENOEXEC;
738 }
739 }
740
741 return 0;
742}
743
744#ifdef CONFIG_DYNAMIC_FTRACE
745int module_trampoline_target(struct module *mod, unsigned long addr,
746 unsigned long *target)
747{
748 struct ppc64_stub_entry *stub;
749 func_desc_t funcdata;
750 u32 magic;
751
752 if (!within_module_core(addr, mod)) {
753 pr_err("%s: stub %lx not in module %s\n", __func__, addr, mod->name);
754 return -EFAULT;
755 }
756
757 stub = (struct ppc64_stub_entry *)addr;
758
759 if (copy_from_kernel_nofault(&magic, &stub->magic,
760 sizeof(magic))) {
761 pr_err("%s: fault reading magic for stub %lx for %s\n", __func__, addr, mod->name);
762 return -EFAULT;
763 }
764
765 if (magic != STUB_MAGIC) {
766 pr_err("%s: bad magic for stub %lx for %s\n", __func__, addr, mod->name);
767 return -EFAULT;
768 }
769
770 if (copy_from_kernel_nofault(&funcdata, &stub->funcdata,
771 sizeof(funcdata))) {
772 pr_err("%s: fault reading funcdata for stub %lx for %s\n", __func__, addr, mod->name);
773 return -EFAULT;
774 }
775
776 *target = stub_func_addr(funcdata);
777
778 return 0;
779}
780
781int module_finalize_ftrace(struct module *mod, const Elf_Shdr *sechdrs)
782{
783 mod->arch.tramp = stub_for_addr(sechdrs,
784 (unsigned long)ftrace_caller,
785 mod,
786 "ftrace_caller");
787#ifdef CONFIG_DYNAMIC_FTRACE_WITH_REGS
788 mod->arch.tramp_regs = stub_for_addr(sechdrs,
789 (unsigned long)ftrace_regs_caller,
790 mod,
791 "ftrace_regs_caller");
792 if (!mod->arch.tramp_regs)
793 return -ENOENT;
794#endif
795
796 if (!mod->arch.tramp)
797 return -ENOENT;
798
799 return 0;
800}
801#endif
1// SPDX-License-Identifier: GPL-2.0-or-later
2/* Kernel module help for PPC64.
3 Copyright (C) 2001, 2003 Rusty Russell IBM Corporation.
4
5*/
6
7#define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
8
9#include <linux/module.h>
10#include <linux/elf.h>
11#include <linux/moduleloader.h>
12#include <linux/err.h>
13#include <linux/vmalloc.h>
14#include <linux/ftrace.h>
15#include <linux/bug.h>
16#include <linux/uaccess.h>
17#include <linux/kernel.h>
18#include <asm/module.h>
19#include <asm/firmware.h>
20#include <asm/code-patching.h>
21#include <linux/sort.h>
22#include <asm/setup.h>
23#include <asm/sections.h>
24#include <asm/inst.h>
25
26/* FIXME: We don't do .init separately. To do this, we'd need to have
27 a separate r2 value in the init and core section, and stub between
28 them, too.
29
30 Using a magic allocator which places modules within 32MB solves
31 this, and makes other things simpler. Anton?
32 --RR. */
33
34bool module_elf_check_arch(Elf_Ehdr *hdr)
35{
36 unsigned long abi_level = hdr->e_flags & 0x3;
37
38 if (IS_ENABLED(CONFIG_PPC64_ELF_ABI_V2))
39 return abi_level == 2;
40 else
41 return abi_level < 2;
42}
43
44#ifdef CONFIG_PPC64_ELF_ABI_V2
45
46static func_desc_t func_desc(unsigned long addr)
47{
48 func_desc_t desc = {
49 .addr = addr,
50 };
51
52 return desc;
53}
54
55/* PowerPC64 specific values for the Elf64_Sym st_other field. */
56#define STO_PPC64_LOCAL_BIT 5
57#define STO_PPC64_LOCAL_MASK (7 << STO_PPC64_LOCAL_BIT)
58#define PPC64_LOCAL_ENTRY_OFFSET(other) \
59 (((1 << (((other) & STO_PPC64_LOCAL_MASK) >> STO_PPC64_LOCAL_BIT)) >> 2) << 2)
60
61static unsigned int local_entry_offset(const Elf64_Sym *sym)
62{
63 /* sym->st_other indicates offset to local entry point
64 * (otherwise it will assume r12 is the address of the start
65 * of function and try to derive r2 from it). */
66 return PPC64_LOCAL_ENTRY_OFFSET(sym->st_other);
67}
68#else
69
70static func_desc_t func_desc(unsigned long addr)
71{
72 return *(struct func_desc *)addr;
73}
74static unsigned int local_entry_offset(const Elf64_Sym *sym)
75{
76 return 0;
77}
78
79void *dereference_module_function_descriptor(struct module *mod, void *ptr)
80{
81 if (ptr < (void *)mod->arch.start_opd ||
82 ptr >= (void *)mod->arch.end_opd)
83 return ptr;
84
85 return dereference_function_descriptor(ptr);
86}
87#endif
88
89static unsigned long func_addr(unsigned long addr)
90{
91 return func_desc(addr).addr;
92}
93
94static unsigned long stub_func_addr(func_desc_t func)
95{
96 return func.addr;
97}
98
99#define STUB_MAGIC 0x73747562 /* stub */
100
101/* Like PPC32, we need little trampolines to do > 24-bit jumps (into
102 the kernel itself). But on PPC64, these need to be used for every
103 jump, actually, to reset r2 (TOC+0x8000). */
104struct ppc64_stub_entry {
105 /*
106 * 28 byte jump instruction sequence (7 instructions) that can
107 * hold ppc64_stub_insns or stub_insns. Must be 8-byte aligned
108 * with PCREL kernels that use prefix instructions in the stub.
109 */
110 u32 jump[7];
111 /* Used by ftrace to identify stubs */
112 u32 magic;
113 /* Data for the above code */
114 func_desc_t funcdata;
115} __aligned(8);
116
117struct ppc64_got_entry {
118 u64 addr;
119};
120
121/*
122 * PPC64 uses 24 bit jumps, but we need to jump into other modules or
123 * the kernel which may be further. So we jump to a stub.
124 *
125 * Target address and TOC are loaded from function descriptor in the
126 * ppc64_stub_entry.
127 *
128 * r12 is used to generate the target address, which is required for the
129 * ELFv2 global entry point calling convention.
130 *
131 * TOC handling:
132 * - PCREL does not have a TOC.
133 * - ELFv2 non-PCREL just has to save r2, the callee is responsible for
134 * setting its own TOC pointer at the global entry address.
135 * - ELFv1 must load the new TOC pointer from the function descriptor.
136 */
137static u32 ppc64_stub_insns[] = {
138#ifdef CONFIG_PPC_KERNEL_PCREL
139 /* pld r12,addr */
140 PPC_PREFIX_8LS | __PPC_PRFX_R(1),
141 PPC_INST_PLD | ___PPC_RT(_R12),
142#else
143 PPC_RAW_ADDIS(_R11, _R2, 0),
144 PPC_RAW_ADDI(_R11, _R11, 0),
145 /* Save current r2 value in magic place on the stack. */
146 PPC_RAW_STD(_R2, _R1, R2_STACK_OFFSET),
147 PPC_RAW_LD(_R12, _R11, 32),
148#ifdef CONFIG_PPC64_ELF_ABI_V1
149 /* Set up new r2 from function descriptor */
150 PPC_RAW_LD(_R2, _R11, 40),
151#endif
152#endif
153 PPC_RAW_MTCTR(_R12),
154 PPC_RAW_BCTR(),
155};
156
157/*
158 * Count how many different r_type relocations (different symbol,
159 * different addend).
160 */
161static unsigned int count_relocs(const Elf64_Rela *rela, unsigned int num,
162 unsigned long r_type)
163{
164 unsigned int i, r_info, r_addend, _count_relocs;
165
166 /* FIXME: Only count external ones --RR */
167 _count_relocs = 0;
168 r_info = 0;
169 r_addend = 0;
170 for (i = 0; i < num; i++)
171 /* Only count r_type relocs, others don't need stubs */
172 if (ELF64_R_TYPE(rela[i].r_info) == r_type &&
173 (r_info != ELF64_R_SYM(rela[i].r_info) ||
174 r_addend != rela[i].r_addend)) {
175 _count_relocs++;
176 r_info = ELF64_R_SYM(rela[i].r_info);
177 r_addend = rela[i].r_addend;
178 }
179
180 return _count_relocs;
181}
182
183static int relacmp(const void *_x, const void *_y)
184{
185 const Elf64_Rela *x, *y;
186
187 y = (Elf64_Rela *)_x;
188 x = (Elf64_Rela *)_y;
189
190 /* Compare the entire r_info (as opposed to ELF64_R_SYM(r_info) only) to
191 * make the comparison cheaper/faster. It won't affect the sorting or
192 * the counting algorithms' performance
193 */
194 if (x->r_info < y->r_info)
195 return -1;
196 else if (x->r_info > y->r_info)
197 return 1;
198 else if (x->r_addend < y->r_addend)
199 return -1;
200 else if (x->r_addend > y->r_addend)
201 return 1;
202 else
203 return 0;
204}
205
206/* Get size of potential trampolines required. */
207static unsigned long get_stubs_size(const Elf64_Ehdr *hdr,
208 const Elf64_Shdr *sechdrs)
209{
210 /* One extra reloc so it's always 0-addr terminated */
211 unsigned long relocs = 1;
212 unsigned i;
213
214 /* Every relocated section... */
215 for (i = 1; i < hdr->e_shnum; i++) {
216 if (sechdrs[i].sh_type == SHT_RELA) {
217 pr_debug("Found relocations in section %u\n", i);
218 pr_debug("Ptr: %p. Number: %Lu\n",
219 (void *)sechdrs[i].sh_addr,
220 sechdrs[i].sh_size / sizeof(Elf64_Rela));
221
222 /* Sort the relocation information based on a symbol and
223 * addend key. This is a stable O(n*log n) complexity
224 * algorithm but it will reduce the complexity of
225 * count_relocs() to linear complexity O(n)
226 */
227 sort((void *)sechdrs[i].sh_addr,
228 sechdrs[i].sh_size / sizeof(Elf64_Rela),
229 sizeof(Elf64_Rela), relacmp, NULL);
230
231 relocs += count_relocs((void *)sechdrs[i].sh_addr,
232 sechdrs[i].sh_size
233 / sizeof(Elf64_Rela),
234 R_PPC_REL24);
235#ifdef CONFIG_PPC_KERNEL_PCREL
236 relocs += count_relocs((void *)sechdrs[i].sh_addr,
237 sechdrs[i].sh_size
238 / sizeof(Elf64_Rela),
239 R_PPC64_REL24_NOTOC);
240#endif
241 }
242 }
243
244#ifdef CONFIG_DYNAMIC_FTRACE
245 /* make the trampoline to the ftrace_caller */
246 relocs++;
247#ifdef CONFIG_DYNAMIC_FTRACE_WITH_REGS
248 /* an additional one for ftrace_regs_caller */
249 relocs++;
250#endif
251#endif
252
253 pr_debug("Looks like a total of %lu stubs, max\n", relocs);
254 return relocs * sizeof(struct ppc64_stub_entry);
255}
256
257#ifdef CONFIG_PPC_KERNEL_PCREL
258static int count_pcpu_relocs(const Elf64_Shdr *sechdrs,
259 const Elf64_Rela *rela, unsigned int num,
260 unsigned int symindex, unsigned int pcpu)
261{
262 unsigned int i, r_info, r_addend, _count_relocs;
263
264 _count_relocs = 0;
265 r_info = 0;
266 r_addend = 0;
267
268 for (i = 0; i < num; i++) {
269 Elf64_Sym *sym;
270
271 /* This is the symbol it is referring to */
272 sym = (Elf64_Sym *)sechdrs[symindex].sh_addr
273 + ELF64_R_SYM(rela[i].r_info);
274
275 if (sym->st_shndx == pcpu &&
276 (r_info != ELF64_R_SYM(rela[i].r_info) ||
277 r_addend != rela[i].r_addend)) {
278 _count_relocs++;
279 r_info = ELF64_R_SYM(rela[i].r_info);
280 r_addend = rela[i].r_addend;
281 }
282 }
283
284 return _count_relocs;
285}
286
287/* Get size of potential GOT required. */
288static unsigned long get_got_size(const Elf64_Ehdr *hdr,
289 const Elf64_Shdr *sechdrs,
290 struct module *me)
291{
292 /* One extra reloc so it's always 0-addr terminated */
293 unsigned long relocs = 1;
294 unsigned int i, symindex = 0;
295
296 for (i = 1; i < hdr->e_shnum; i++) {
297 if (sechdrs[i].sh_type == SHT_SYMTAB) {
298 symindex = i;
299 break;
300 }
301 }
302 WARN_ON_ONCE(!symindex);
303
304 /* Every relocated section... */
305 for (i = 1; i < hdr->e_shnum; i++) {
306 if (sechdrs[i].sh_type == SHT_RELA) {
307 pr_debug("Found relocations in section %u\n", i);
308 pr_debug("Ptr: %p. Number: %llu\n", (void *)sechdrs[i].sh_addr,
309 sechdrs[i].sh_size / sizeof(Elf64_Rela));
310
311 /*
312 * Sort the relocation information based on a symbol and
313 * addend key. This is a stable O(n*log n) complexity
314 * algorithm but it will reduce the complexity of
315 * count_relocs() to linear complexity O(n)
316 */
317 sort((void *)sechdrs[i].sh_addr,
318 sechdrs[i].sh_size / sizeof(Elf64_Rela),
319 sizeof(Elf64_Rela), relacmp, NULL);
320
321 relocs += count_relocs((void *)sechdrs[i].sh_addr,
322 sechdrs[i].sh_size
323 / sizeof(Elf64_Rela),
324 R_PPC64_GOT_PCREL34);
325
326 /*
327 * Percpu data access typically gets linked with
328 * REL34 relocations, but the percpu section gets
329 * moved at load time and requires that to be
330 * converted to GOT linkage.
331 */
332 if (IS_ENABLED(CONFIG_SMP) && symindex)
333 relocs += count_pcpu_relocs(sechdrs,
334 (void *)sechdrs[i].sh_addr,
335 sechdrs[i].sh_size
336 / sizeof(Elf64_Rela),
337 symindex, me->arch.pcpu_section);
338 }
339 }
340
341 pr_debug("Looks like a total of %lu GOT entries, max\n", relocs);
342 return relocs * sizeof(struct ppc64_got_entry);
343}
344#else /* CONFIG_PPC_KERNEL_PCREL */
345
346/* Still needed for ELFv2, for .TOC. */
347static void dedotify_versions(struct modversion_info *vers,
348 unsigned long size)
349{
350 struct modversion_info *end;
351
352 for (end = (void *)vers + size; vers < end; vers++)
353 if (vers->name[0] == '.') {
354 memmove(vers->name, vers->name+1, strlen(vers->name));
355 }
356}
357
358/*
359 * Undefined symbols which refer to .funcname, hack to funcname. Make .TOC.
360 * seem to be defined (value set later).
361 */
362static void dedotify(Elf64_Sym *syms, unsigned int numsyms, char *strtab)
363{
364 unsigned int i;
365
366 for (i = 1; i < numsyms; i++) {
367 if (syms[i].st_shndx == SHN_UNDEF) {
368 char *name = strtab + syms[i].st_name;
369 if (name[0] == '.') {
370 if (strcmp(name+1, "TOC.") == 0)
371 syms[i].st_shndx = SHN_ABS;
372 syms[i].st_name++;
373 }
374 }
375 }
376}
377
378static Elf64_Sym *find_dot_toc(Elf64_Shdr *sechdrs,
379 const char *strtab,
380 unsigned int symindex)
381{
382 unsigned int i, numsyms;
383 Elf64_Sym *syms;
384
385 syms = (Elf64_Sym *)sechdrs[symindex].sh_addr;
386 numsyms = sechdrs[symindex].sh_size / sizeof(Elf64_Sym);
387
388 for (i = 1; i < numsyms; i++) {
389 if (syms[i].st_shndx == SHN_ABS
390 && strcmp(strtab + syms[i].st_name, "TOC.") == 0)
391 return &syms[i];
392 }
393 return NULL;
394}
395#endif /* CONFIG_PPC_KERNEL_PCREL */
396
397bool module_init_section(const char *name)
398{
399 /* We don't handle .init for the moment: always return false. */
400 return false;
401}
402
403int module_frob_arch_sections(Elf64_Ehdr *hdr,
404 Elf64_Shdr *sechdrs,
405 char *secstrings,
406 struct module *me)
407{
408 unsigned int i;
409
410 /* Find .toc and .stubs sections, symtab and strtab */
411 for (i = 1; i < hdr->e_shnum; i++) {
412 if (strcmp(secstrings + sechdrs[i].sh_name, ".stubs") == 0)
413 me->arch.stubs_section = i;
414#ifdef CONFIG_PPC_KERNEL_PCREL
415 else if (strcmp(secstrings + sechdrs[i].sh_name, ".data..percpu") == 0)
416 me->arch.pcpu_section = i;
417 else if (strcmp(secstrings + sechdrs[i].sh_name, ".mygot") == 0) {
418 me->arch.got_section = i;
419 if (sechdrs[i].sh_addralign < 8)
420 sechdrs[i].sh_addralign = 8;
421 }
422#else
423 else if (strcmp(secstrings + sechdrs[i].sh_name, ".toc") == 0) {
424 me->arch.toc_section = i;
425 if (sechdrs[i].sh_addralign < 8)
426 sechdrs[i].sh_addralign = 8;
427 }
428 else if (strcmp(secstrings+sechdrs[i].sh_name,"__versions")==0)
429 dedotify_versions((void *)hdr + sechdrs[i].sh_offset,
430 sechdrs[i].sh_size);
431
432 if (sechdrs[i].sh_type == SHT_SYMTAB)
433 dedotify((void *)hdr + sechdrs[i].sh_offset,
434 sechdrs[i].sh_size / sizeof(Elf64_Sym),
435 (void *)hdr
436 + sechdrs[sechdrs[i].sh_link].sh_offset);
437#endif
438 }
439
440 if (!me->arch.stubs_section) {
441 pr_err("%s: doesn't contain .stubs.\n", me->name);
442 return -ENOEXEC;
443 }
444
445#ifdef CONFIG_PPC_KERNEL_PCREL
446 if (!me->arch.got_section) {
447 pr_err("%s: doesn't contain .mygot.\n", me->name);
448 return -ENOEXEC;
449 }
450
451 /* Override the got size */
452 sechdrs[me->arch.got_section].sh_size = get_got_size(hdr, sechdrs, me);
453#else
454 /* If we don't have a .toc, just use .stubs. We need to set r2
455 to some reasonable value in case the module calls out to
456 other functions via a stub, or if a function pointer escapes
457 the module by some means. */
458 if (!me->arch.toc_section)
459 me->arch.toc_section = me->arch.stubs_section;
460#endif
461
462 /* Override the stubs size */
463 sechdrs[me->arch.stubs_section].sh_size = get_stubs_size(hdr, sechdrs);
464
465 return 0;
466}
467
468#if defined(CONFIG_MPROFILE_KERNEL) || defined(CONFIG_ARCH_USING_PATCHABLE_FUNCTION_ENTRY)
469
470static u32 stub_insns[] = {
471#ifdef CONFIG_PPC_KERNEL_PCREL
472 PPC_RAW_LD(_R12, _R13, offsetof(struct paca_struct, kernelbase)),
473 PPC_RAW_NOP(), /* align the prefix insn */
474 /* paddi r12,r12,addr */
475 PPC_PREFIX_MLS | __PPC_PRFX_R(0),
476 PPC_INST_PADDI | ___PPC_RT(_R12) | ___PPC_RA(_R12),
477 PPC_RAW_MTCTR(_R12),
478 PPC_RAW_BCTR(),
479#else
480 PPC_RAW_LD(_R12, _R13, offsetof(struct paca_struct, kernel_toc)),
481 PPC_RAW_ADDIS(_R12, _R12, 0),
482 PPC_RAW_ADDI(_R12, _R12, 0),
483 PPC_RAW_MTCTR(_R12),
484 PPC_RAW_BCTR(),
485#endif
486};
487
488/*
489 * For mprofile-kernel we use a special stub for ftrace_caller() because we
490 * can't rely on r2 containing this module's TOC when we enter the stub.
491 *
492 * That can happen if the function calling us didn't need to use the toc. In
493 * that case it won't have setup r2, and the r2 value will be either the
494 * kernel's toc, or possibly another modules toc.
495 *
496 * To deal with that this stub uses the kernel toc, which is always accessible
497 * via the paca (in r13). The target (ftrace_caller()) is responsible for
498 * saving and restoring the toc before returning.
499 */
500static inline int create_ftrace_stub(struct ppc64_stub_entry *entry,
501 unsigned long addr,
502 struct module *me)
503{
504 long reladdr;
505
506 if ((unsigned long)entry->jump % 8 != 0) {
507 pr_err("%s: Address of stub entry is not 8-byte aligned\n", me->name);
508 return 0;
509 }
510
511 BUILD_BUG_ON(sizeof(stub_insns) > sizeof(entry->jump));
512 memcpy(entry->jump, stub_insns, sizeof(stub_insns));
513
514 if (IS_ENABLED(CONFIG_PPC_KERNEL_PCREL)) {
515 /* Stub uses address relative to kernel base (from the paca) */
516 reladdr = addr - local_paca->kernelbase;
517 if (reladdr > 0x1FFFFFFFFL || reladdr < -0x200000000L) {
518 pr_err("%s: Address of %ps out of range of 34-bit relative address.\n",
519 me->name, (void *)addr);
520 return 0;
521 }
522
523 entry->jump[2] |= IMM_H18(reladdr);
524 entry->jump[3] |= IMM_L(reladdr);
525 } else {
526 /* Stub uses address relative to kernel toc (from the paca) */
527 reladdr = addr - kernel_toc_addr();
528 if (reladdr > 0x7FFFFFFF || reladdr < -(0x80000000L)) {
529 pr_err("%s: Address of %ps out of range of kernel_toc.\n",
530 me->name, (void *)addr);
531 return 0;
532 }
533
534 entry->jump[1] |= PPC_HA(reladdr);
535 entry->jump[2] |= PPC_LO(reladdr);
536 }
537
538 /* Even though we don't use funcdata in the stub, it's needed elsewhere. */
539 entry->funcdata = func_desc(addr);
540 entry->magic = STUB_MAGIC;
541
542 return 1;
543}
544
545static bool is_mprofile_ftrace_call(const char *name)
546{
547 if (!strcmp("_mcount", name))
548 return true;
549#ifdef CONFIG_DYNAMIC_FTRACE
550 if (!strcmp("ftrace_caller", name))
551 return true;
552#ifdef CONFIG_DYNAMIC_FTRACE_WITH_REGS
553 if (!strcmp("ftrace_regs_caller", name))
554 return true;
555#endif
556#endif
557
558 return false;
559}
560#else
561static inline int create_ftrace_stub(struct ppc64_stub_entry *entry,
562 unsigned long addr,
563 struct module *me)
564{
565 return 0;
566}
567
568static bool is_mprofile_ftrace_call(const char *name)
569{
570 return false;
571}
572#endif
573
574/*
575 * r2 is the TOC pointer: it actually points 0x8000 into the TOC (this gives the
576 * value maximum span in an instruction which uses a signed offset). Round down
577 * to a 256 byte boundary for the odd case where we are setting up r2 without a
578 * .toc section.
579 */
580static inline unsigned long my_r2(const Elf64_Shdr *sechdrs, struct module *me)
581{
582#ifndef CONFIG_PPC_KERNEL_PCREL
583 return (sechdrs[me->arch.toc_section].sh_addr & ~0xfful) + 0x8000;
584#else
585 return -1;
586#endif
587}
588
589/* Patch stub to reference function and correct r2 value. */
590static inline int create_stub(const Elf64_Shdr *sechdrs,
591 struct ppc64_stub_entry *entry,
592 unsigned long addr,
593 struct module *me,
594 const char *name)
595{
596 long reladdr;
597 func_desc_t desc;
598 int i;
599
600 if (is_mprofile_ftrace_call(name))
601 return create_ftrace_stub(entry, addr, me);
602
603 if ((unsigned long)entry->jump % 8 != 0) {
604 pr_err("%s: Address of stub entry is not 8-byte aligned\n", me->name);
605 return 0;
606 }
607
608 BUILD_BUG_ON(sizeof(ppc64_stub_insns) > sizeof(entry->jump));
609 for (i = 0; i < ARRAY_SIZE(ppc64_stub_insns); i++) {
610 if (patch_instruction(&entry->jump[i],
611 ppc_inst(ppc64_stub_insns[i])))
612 return 0;
613 }
614
615 if (IS_ENABLED(CONFIG_PPC_KERNEL_PCREL)) {
616 /* Stub uses address relative to itself! */
617 reladdr = 0 + offsetof(struct ppc64_stub_entry, funcdata);
618 BUILD_BUG_ON(reladdr != 32);
619 if (reladdr > 0x1FFFFFFFFL || reladdr < -0x200000000L) {
620 pr_err("%s: Address of %p out of range of 34-bit relative address.\n",
621 me->name, (void *)reladdr);
622 return 0;
623 }
624 pr_debug("Stub %p get data from reladdr %li\n", entry, reladdr);
625
626 /* May not even need this if we're relative to 0 */
627 if (patch_instruction(&entry->jump[0],
628 ppc_inst_prefix(entry->jump[0] | IMM_H18(reladdr),
629 entry->jump[1] | IMM_L(reladdr))))
630 return 0;
631
632 } else {
633 /* Stub uses address relative to r2. */
634 reladdr = (unsigned long)entry - my_r2(sechdrs, me);
635 if (reladdr > 0x7FFFFFFF || reladdr < -(0x80000000L)) {
636 pr_err("%s: Address %p of stub out of range of %p.\n",
637 me->name, (void *)reladdr, (void *)my_r2);
638 return 0;
639 }
640 pr_debug("Stub %p get data from reladdr %li\n", entry, reladdr);
641
642 if (patch_instruction(&entry->jump[0],
643 ppc_inst(entry->jump[0] | PPC_HA(reladdr))))
644 return 0;
645
646 if (patch_instruction(&entry->jump[1],
647 ppc_inst(entry->jump[1] | PPC_LO(reladdr))))
648 return 0;
649 }
650
651 // func_desc_t is 8 bytes if ABIv2, else 16 bytes
652 desc = func_desc(addr);
653 for (i = 0; i < sizeof(func_desc_t) / sizeof(u32); i++) {
654 if (patch_instruction(((u32 *)&entry->funcdata) + i,
655 ppc_inst(((u32 *)(&desc))[i])))
656 return 0;
657 }
658
659 if (patch_instruction(&entry->magic, ppc_inst(STUB_MAGIC)))
660 return 0;
661
662 return 1;
663}
664
665/* Create stub to jump to function described in this OPD/ptr: we need the
666 stub to set up the TOC ptr (r2) for the function. */
667static unsigned long stub_for_addr(const Elf64_Shdr *sechdrs,
668 unsigned long addr,
669 struct module *me,
670 const char *name)
671{
672 struct ppc64_stub_entry *stubs;
673 unsigned int i, num_stubs;
674
675 num_stubs = sechdrs[me->arch.stubs_section].sh_size / sizeof(*stubs);
676
677 /* Find this stub, or if that fails, the next avail. entry */
678 stubs = (void *)sechdrs[me->arch.stubs_section].sh_addr;
679 for (i = 0; stub_func_addr(stubs[i].funcdata); i++) {
680 if (WARN_ON(i >= num_stubs))
681 return 0;
682
683 if (stub_func_addr(stubs[i].funcdata) == func_addr(addr))
684 return (unsigned long)&stubs[i];
685 }
686
687 if (!create_stub(sechdrs, &stubs[i], addr, me, name))
688 return 0;
689
690 return (unsigned long)&stubs[i];
691}
692
693#ifdef CONFIG_PPC_KERNEL_PCREL
694/* Create GOT to load the location described in this ptr */
695static unsigned long got_for_addr(const Elf64_Shdr *sechdrs,
696 unsigned long addr,
697 struct module *me,
698 const char *name)
699{
700 struct ppc64_got_entry *got;
701 unsigned int i, num_got;
702
703 if (!IS_ENABLED(CONFIG_PPC_KERNEL_PCREL))
704 return addr;
705
706 num_got = sechdrs[me->arch.got_section].sh_size / sizeof(*got);
707
708 /* Find this stub, or if that fails, the next avail. entry */
709 got = (void *)sechdrs[me->arch.got_section].sh_addr;
710 for (i = 0; got[i].addr; i++) {
711 if (WARN_ON(i >= num_got))
712 return 0;
713
714 if (got[i].addr == addr)
715 return (unsigned long)&got[i];
716 }
717
718 got[i].addr = addr;
719
720 return (unsigned long)&got[i];
721}
722#endif
723
724/* We expect a noop next: if it is, replace it with instruction to
725 restore r2. */
726static int restore_r2(const char *name, u32 *instruction, struct module *me)
727{
728 u32 *prev_insn = instruction - 1;
729 u32 insn_val = *instruction;
730
731 if (IS_ENABLED(CONFIG_PPC_KERNEL_PCREL))
732 return 0;
733
734 if (is_mprofile_ftrace_call(name))
735 return 0;
736
737 /*
738 * Make sure the branch isn't a sibling call. Sibling calls aren't
739 * "link" branches and they don't return, so they don't need the r2
740 * restore afterwards.
741 */
742 if (!instr_is_relative_link_branch(ppc_inst(*prev_insn)))
743 return 0;
744
745 /*
746 * For livepatch, the restore r2 instruction might have already been
747 * written previously, if the referenced symbol is in a previously
748 * unloaded module which is now being loaded again. In that case, skip
749 * the warning and the instruction write.
750 */
751 if (insn_val == PPC_INST_LD_TOC)
752 return 0;
753
754 if (insn_val != PPC_RAW_NOP()) {
755 pr_err("%s: Expected nop after call, got %08x at %pS\n",
756 me->name, insn_val, instruction);
757 return -ENOEXEC;
758 }
759
760 /* ld r2,R2_STACK_OFFSET(r1) */
761 return patch_instruction(instruction, ppc_inst(PPC_INST_LD_TOC));
762}
763
764int apply_relocate_add(Elf64_Shdr *sechdrs,
765 const char *strtab,
766 unsigned int symindex,
767 unsigned int relsec,
768 struct module *me)
769{
770 unsigned int i;
771 Elf64_Rela *rela = (void *)sechdrs[relsec].sh_addr;
772 Elf64_Sym *sym;
773 unsigned long *location;
774 unsigned long value;
775
776 pr_debug("Applying ADD relocate section %u to %u\n", relsec,
777 sechdrs[relsec].sh_info);
778
779#ifndef CONFIG_PPC_KERNEL_PCREL
780 /* First time we're called, we can fix up .TOC. */
781 if (!me->arch.toc_fixed) {
782 sym = find_dot_toc(sechdrs, strtab, symindex);
783 /* It's theoretically possible that a module doesn't want a
784 * .TOC. so don't fail it just for that. */
785 if (sym)
786 sym->st_value = my_r2(sechdrs, me);
787 me->arch.toc_fixed = true;
788 }
789#endif
790 for (i = 0; i < sechdrs[relsec].sh_size / sizeof(*rela); i++) {
791 /* This is where to make the change */
792 location = (void *)sechdrs[sechdrs[relsec].sh_info].sh_addr
793 + rela[i].r_offset;
794 /* This is the symbol it is referring to */
795 sym = (Elf64_Sym *)sechdrs[symindex].sh_addr
796 + ELF64_R_SYM(rela[i].r_info);
797
798 pr_debug("RELOC at %p: %li-type as %s (0x%lx) + %li\n",
799 location, (long)ELF64_R_TYPE(rela[i].r_info),
800 strtab + sym->st_name, (unsigned long)sym->st_value,
801 (long)rela[i].r_addend);
802
803 /* `Everything is relative'. */
804 value = sym->st_value + rela[i].r_addend;
805
806 switch (ELF64_R_TYPE(rela[i].r_info)) {
807 case R_PPC64_ADDR32:
808 /* Simply set it */
809 *(u32 *)location = value;
810 break;
811
812 case R_PPC64_ADDR64:
813 /* Simply set it */
814 *(unsigned long *)location = value;
815 break;
816
817#ifndef CONFIG_PPC_KERNEL_PCREL
818 case R_PPC64_TOC:
819 *(unsigned long *)location = my_r2(sechdrs, me);
820 break;
821
822 case R_PPC64_TOC16:
823 /* Subtract TOC pointer */
824 value -= my_r2(sechdrs, me);
825 if (value + 0x8000 > 0xffff) {
826 pr_err("%s: bad TOC16 relocation (0x%lx)\n",
827 me->name, value);
828 return -ENOEXEC;
829 }
830 *((uint16_t *) location)
831 = (*((uint16_t *) location) & ~0xffff)
832 | (value & 0xffff);
833 break;
834
835 case R_PPC64_TOC16_LO:
836 /* Subtract TOC pointer */
837 value -= my_r2(sechdrs, me);
838 *((uint16_t *) location)
839 = (*((uint16_t *) location) & ~0xffff)
840 | (value & 0xffff);
841 break;
842
843 case R_PPC64_TOC16_DS:
844 /* Subtract TOC pointer */
845 value -= my_r2(sechdrs, me);
846 if ((value & 3) != 0 || value + 0x8000 > 0xffff) {
847 pr_err("%s: bad TOC16_DS relocation (0x%lx)\n",
848 me->name, value);
849 return -ENOEXEC;
850 }
851 *((uint16_t *) location)
852 = (*((uint16_t *) location) & ~0xfffc)
853 | (value & 0xfffc);
854 break;
855
856 case R_PPC64_TOC16_LO_DS:
857 /* Subtract TOC pointer */
858 value -= my_r2(sechdrs, me);
859 if ((value & 3) != 0) {
860 pr_err("%s: bad TOC16_LO_DS relocation (0x%lx)\n",
861 me->name, value);
862 return -ENOEXEC;
863 }
864 *((uint16_t *) location)
865 = (*((uint16_t *) location) & ~0xfffc)
866 | (value & 0xfffc);
867 break;
868
869 case R_PPC64_TOC16_HA:
870 /* Subtract TOC pointer */
871 value -= my_r2(sechdrs, me);
872 value = ((value + 0x8000) >> 16);
873 *((uint16_t *) location)
874 = (*((uint16_t *) location) & ~0xffff)
875 | (value & 0xffff);
876 break;
877#endif
878
879 case R_PPC_REL24:
880#ifdef CONFIG_PPC_KERNEL_PCREL
881 /* PCREL still generates REL24 for mcount */
882 case R_PPC64_REL24_NOTOC:
883#endif
884 /* FIXME: Handle weak symbols here --RR */
885 if (sym->st_shndx == SHN_UNDEF ||
886 sym->st_shndx == SHN_LIVEPATCH) {
887 /* External: go via stub */
888 value = stub_for_addr(sechdrs, value, me,
889 strtab + sym->st_name);
890 if (!value)
891 return -ENOENT;
892 if (restore_r2(strtab + sym->st_name,
893 (u32 *)location + 1, me))
894 return -ENOEXEC;
895 } else
896 value += local_entry_offset(sym);
897
898 /* Convert value to relative */
899 value -= (unsigned long)location;
900 if (value + 0x2000000 > 0x3ffffff || (value & 3) != 0){
901 pr_err("%s: REL24 %li out of range!\n",
902 me->name, (long int)value);
903 return -ENOEXEC;
904 }
905
906 /* Only replace bits 2 through 26 */
907 value = (*(uint32_t *)location & ~PPC_LI_MASK) | PPC_LI(value);
908
909 if (patch_instruction((u32 *)location, ppc_inst(value)))
910 return -EFAULT;
911
912 break;
913
914 case R_PPC64_REL64:
915 /* 64 bits relative (used by features fixups) */
916 *location = value - (unsigned long)location;
917 break;
918
919 case R_PPC64_REL32:
920 /* 32 bits relative (used by relative exception tables) */
921 /* Convert value to relative */
922 value -= (unsigned long)location;
923 if (value + 0x80000000 > 0xffffffff) {
924 pr_err("%s: REL32 %li out of range!\n",
925 me->name, (long int)value);
926 return -ENOEXEC;
927 }
928 *(u32 *)location = value;
929 break;
930
931#ifdef CONFIG_PPC_KERNEL_PCREL
932 case R_PPC64_PCREL34: {
933 unsigned long absvalue = value;
934
935 /* Convert value to relative */
936 value -= (unsigned long)location;
937
938 if (value + 0x200000000 > 0x3ffffffff) {
939 if (sym->st_shndx != me->arch.pcpu_section) {
940 pr_err("%s: REL34 %li out of range!\n",
941 me->name, (long)value);
942 return -ENOEXEC;
943 }
944
945 /*
946 * per-cpu section is special cased because
947 * it is moved during loading, so has to be
948 * converted to use GOT.
949 */
950 value = got_for_addr(sechdrs, absvalue, me,
951 strtab + sym->st_name);
952 if (!value)
953 return -ENOENT;
954 value -= (unsigned long)location;
955
956 /* Turn pla into pld */
957 if (patch_instruction((u32 *)location,
958 ppc_inst_prefix((*(u32 *)location & ~0x02000000),
959 (*((u32 *)location + 1) & ~0xf8000000) | 0xe4000000)))
960 return -EFAULT;
961 }
962
963 if (patch_instruction((u32 *)location,
964 ppc_inst_prefix((*(u32 *)location & ~0x3ffff) | IMM_H18(value),
965 (*((u32 *)location + 1) & ~0xffff) | IMM_L(value))))
966 return -EFAULT;
967
968 break;
969 }
970
971#else
972 case R_PPC64_TOCSAVE:
973 /*
974 * Marker reloc indicates we don't have to save r2.
975 * That would only save us one instruction, so ignore
976 * it.
977 */
978 break;
979#endif
980
981 case R_PPC64_ENTRY:
982 if (IS_ENABLED(CONFIG_PPC_KERNEL_PCREL))
983 break;
984
985 /*
986 * Optimize ELFv2 large code model entry point if
987 * the TOC is within 2GB range of current location.
988 */
989 value = my_r2(sechdrs, me) - (unsigned long)location;
990 if (value + 0x80008000 > 0xffffffff)
991 break;
992 /*
993 * Check for the large code model prolog sequence:
994 * ld r2, ...(r12)
995 * add r2, r2, r12
996 */
997 if ((((uint32_t *)location)[0] & ~0xfffc) != PPC_RAW_LD(_R2, _R12, 0))
998 break;
999 if (((uint32_t *)location)[1] != PPC_RAW_ADD(_R2, _R2, _R12))
1000 break;
1001 /*
1002 * If found, replace it with:
1003 * addis r2, r12, (.TOC.-func)@ha
1004 * addi r2, r2, (.TOC.-func)@l
1005 */
1006 ((uint32_t *)location)[0] = PPC_RAW_ADDIS(_R2, _R12, PPC_HA(value));
1007 ((uint32_t *)location)[1] = PPC_RAW_ADDI(_R2, _R2, PPC_LO(value));
1008 break;
1009
1010 case R_PPC64_REL16_HA:
1011 /* Subtract location pointer */
1012 value -= (unsigned long)location;
1013 value = ((value + 0x8000) >> 16);
1014 *((uint16_t *) location)
1015 = (*((uint16_t *) location) & ~0xffff)
1016 | (value & 0xffff);
1017 break;
1018
1019 case R_PPC64_REL16_LO:
1020 /* Subtract location pointer */
1021 value -= (unsigned long)location;
1022 *((uint16_t *) location)
1023 = (*((uint16_t *) location) & ~0xffff)
1024 | (value & 0xffff);
1025 break;
1026
1027#ifdef CONFIG_PPC_KERNEL_PCREL
1028 case R_PPC64_GOT_PCREL34:
1029 value = got_for_addr(sechdrs, value, me,
1030 strtab + sym->st_name);
1031 if (!value)
1032 return -ENOENT;
1033 value -= (unsigned long)location;
1034 ((uint32_t *)location)[0] = (((uint32_t *)location)[0] & ~0x3ffff) |
1035 ((value >> 16) & 0x3ffff);
1036 ((uint32_t *)location)[1] = (((uint32_t *)location)[1] & ~0xffff) |
1037 (value & 0xffff);
1038 break;
1039#endif
1040
1041 default:
1042 pr_err("%s: Unknown ADD relocation: %lu\n",
1043 me->name,
1044 (unsigned long)ELF64_R_TYPE(rela[i].r_info));
1045 return -ENOEXEC;
1046 }
1047 }
1048
1049 return 0;
1050}
1051
1052#ifdef CONFIG_DYNAMIC_FTRACE
1053int module_trampoline_target(struct module *mod, unsigned long addr,
1054 unsigned long *target)
1055{
1056 struct ppc64_stub_entry *stub;
1057 func_desc_t funcdata;
1058 u32 magic;
1059
1060 if (!within_module_core(addr, mod)) {
1061 pr_err("%s: stub %lx not in module %s\n", __func__, addr, mod->name);
1062 return -EFAULT;
1063 }
1064
1065 stub = (struct ppc64_stub_entry *)addr;
1066
1067 if (copy_from_kernel_nofault(&magic, &stub->magic,
1068 sizeof(magic))) {
1069 pr_err("%s: fault reading magic for stub %lx for %s\n", __func__, addr, mod->name);
1070 return -EFAULT;
1071 }
1072
1073 if (magic != STUB_MAGIC) {
1074 pr_err("%s: bad magic for stub %lx for %s\n", __func__, addr, mod->name);
1075 return -EFAULT;
1076 }
1077
1078 if (copy_from_kernel_nofault(&funcdata, &stub->funcdata,
1079 sizeof(funcdata))) {
1080 pr_err("%s: fault reading funcdata for stub %lx for %s\n", __func__, addr, mod->name);
1081 return -EFAULT;
1082 }
1083
1084 *target = stub_func_addr(funcdata);
1085
1086 return 0;
1087}
1088
1089int module_finalize_ftrace(struct module *mod, const Elf_Shdr *sechdrs)
1090{
1091 mod->arch.tramp = stub_for_addr(sechdrs,
1092 (unsigned long)ftrace_caller,
1093 mod,
1094 "ftrace_caller");
1095#ifdef CONFIG_DYNAMIC_FTRACE_WITH_REGS
1096 mod->arch.tramp_regs = stub_for_addr(sechdrs,
1097 (unsigned long)ftrace_regs_caller,
1098 mod,
1099 "ftrace_regs_caller");
1100 if (!mod->arch.tramp_regs)
1101 return -ENOENT;
1102#endif
1103
1104 if (!mod->arch.tramp)
1105 return -ENOENT;
1106
1107 return 0;
1108}
1109#endif