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 <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). We only
107 * need 6 instructions on ABIv2 but we always allocate 7 so
108 * so we don't have to modify the trampoline load instruction. */
109 u32 jump[7];
110 /* Used by ftrace to identify stubs */
111 u32 magic;
112 /* Data for the above code */
113 func_desc_t funcdata;
114};
115
116/*
117 * PPC64 uses 24 bit jumps, but we need to jump into other modules or
118 * the kernel which may be further. So we jump to a stub.
119 *
120 * For ELFv1 we need to use this to set up the new r2 value (aka TOC
121 * pointer). For ELFv2 it's the callee's responsibility to set up the
122 * new r2, but for both we need to save the old r2.
123 *
124 * We could simply patch the new r2 value and function pointer into
125 * the stub, but it's significantly shorter to put these values at the
126 * end of the stub code, and patch the stub address (32-bits relative
127 * to the TOC ptr, r2) into the stub.
128 */
129static u32 ppc64_stub_insns[] = {
130 PPC_RAW_ADDIS(_R11, _R2, 0),
131 PPC_RAW_ADDI(_R11, _R11, 0),
132 /* Save current r2 value in magic place on the stack. */
133 PPC_RAW_STD(_R2, _R1, R2_STACK_OFFSET),
134 PPC_RAW_LD(_R12, _R11, 32),
135#ifdef CONFIG_PPC64_ELF_ABI_V1
136 /* Set up new r2 from function descriptor */
137 PPC_RAW_LD(_R2, _R11, 40),
138#endif
139 PPC_RAW_MTCTR(_R12),
140 PPC_RAW_BCTR(),
141};
142
143/* Count how many different 24-bit relocations (different symbol,
144 different addend) */
145static unsigned int count_relocs(const Elf64_Rela *rela, unsigned int num)
146{
147 unsigned int i, r_info, r_addend, _count_relocs;
148
149 /* FIXME: Only count external ones --RR */
150 _count_relocs = 0;
151 r_info = 0;
152 r_addend = 0;
153 for (i = 0; i < num; i++)
154 /* Only count 24-bit relocs, others don't need stubs */
155 if (ELF64_R_TYPE(rela[i].r_info) == R_PPC_REL24 &&
156 (r_info != ELF64_R_SYM(rela[i].r_info) ||
157 r_addend != rela[i].r_addend)) {
158 _count_relocs++;
159 r_info = ELF64_R_SYM(rela[i].r_info);
160 r_addend = rela[i].r_addend;
161 }
162
163 return _count_relocs;
164}
165
166static int relacmp(const void *_x, const void *_y)
167{
168 const Elf64_Rela *x, *y;
169
170 y = (Elf64_Rela *)_x;
171 x = (Elf64_Rela *)_y;
172
173 /* Compare the entire r_info (as opposed to ELF64_R_SYM(r_info) only) to
174 * make the comparison cheaper/faster. It won't affect the sorting or
175 * the counting algorithms' performance
176 */
177 if (x->r_info < y->r_info)
178 return -1;
179 else if (x->r_info > y->r_info)
180 return 1;
181 else if (x->r_addend < y->r_addend)
182 return -1;
183 else if (x->r_addend > y->r_addend)
184 return 1;
185 else
186 return 0;
187}
188
189/* Get size of potential trampolines required. */
190static unsigned long get_stubs_size(const Elf64_Ehdr *hdr,
191 const Elf64_Shdr *sechdrs)
192{
193 /* One extra reloc so it's always 0-addr terminated */
194 unsigned long relocs = 1;
195 unsigned i;
196
197 /* Every relocated section... */
198 for (i = 1; i < hdr->e_shnum; i++) {
199 if (sechdrs[i].sh_type == SHT_RELA) {
200 pr_debug("Found relocations in section %u\n", i);
201 pr_debug("Ptr: %p. Number: %Lu\n",
202 (void *)sechdrs[i].sh_addr,
203 sechdrs[i].sh_size / sizeof(Elf64_Rela));
204
205 /* Sort the relocation information based on a symbol and
206 * addend key. This is a stable O(n*log n) complexity
207 * algorithm but it will reduce the complexity of
208 * count_relocs() to linear complexity O(n)
209 */
210 sort((void *)sechdrs[i].sh_addr,
211 sechdrs[i].sh_size / sizeof(Elf64_Rela),
212 sizeof(Elf64_Rela), relacmp, NULL);
213
214 relocs += count_relocs((void *)sechdrs[i].sh_addr,
215 sechdrs[i].sh_size
216 / sizeof(Elf64_Rela));
217 }
218 }
219
220#ifdef CONFIG_DYNAMIC_FTRACE
221 /* make the trampoline to the ftrace_caller */
222 relocs++;
223#ifdef CONFIG_DYNAMIC_FTRACE_WITH_REGS
224 /* an additional one for ftrace_regs_caller */
225 relocs++;
226#endif
227#endif
228
229 pr_debug("Looks like a total of %lu stubs, max\n", relocs);
230 return relocs * sizeof(struct ppc64_stub_entry);
231}
232
233/* Still needed for ELFv2, for .TOC. */
234static void dedotify_versions(struct modversion_info *vers,
235 unsigned long size)
236{
237 struct modversion_info *end;
238
239 for (end = (void *)vers + size; vers < end; vers++)
240 if (vers->name[0] == '.') {
241 memmove(vers->name, vers->name+1, strlen(vers->name));
242 }
243}
244
245/*
246 * Undefined symbols which refer to .funcname, hack to funcname. Make .TOC.
247 * seem to be defined (value set later).
248 */
249static void dedotify(Elf64_Sym *syms, unsigned int numsyms, char *strtab)
250{
251 unsigned int i;
252
253 for (i = 1; i < numsyms; i++) {
254 if (syms[i].st_shndx == SHN_UNDEF) {
255 char *name = strtab + syms[i].st_name;
256 if (name[0] == '.') {
257 if (strcmp(name+1, "TOC.") == 0)
258 syms[i].st_shndx = SHN_ABS;
259 syms[i].st_name++;
260 }
261 }
262 }
263}
264
265static Elf64_Sym *find_dot_toc(Elf64_Shdr *sechdrs,
266 const char *strtab,
267 unsigned int symindex)
268{
269 unsigned int i, numsyms;
270 Elf64_Sym *syms;
271
272 syms = (Elf64_Sym *)sechdrs[symindex].sh_addr;
273 numsyms = sechdrs[symindex].sh_size / sizeof(Elf64_Sym);
274
275 for (i = 1; i < numsyms; i++) {
276 if (syms[i].st_shndx == SHN_ABS
277 && strcmp(strtab + syms[i].st_name, "TOC.") == 0)
278 return &syms[i];
279 }
280 return NULL;
281}
282
283bool module_init_section(const char *name)
284{
285 /* We don't handle .init for the moment: always return false. */
286 return false;
287}
288
289int module_frob_arch_sections(Elf64_Ehdr *hdr,
290 Elf64_Shdr *sechdrs,
291 char *secstrings,
292 struct module *me)
293{
294 unsigned int i;
295
296 /* Find .toc and .stubs sections, symtab and strtab */
297 for (i = 1; i < hdr->e_shnum; i++) {
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 if (sechdrs[i].sh_type == SHT_SYMTAB)
310 dedotify((void *)hdr + sechdrs[i].sh_offset,
311 sechdrs[i].sh_size / sizeof(Elf64_Sym),
312 (void *)hdr
313 + sechdrs[sechdrs[i].sh_link].sh_offset);
314 }
315
316 if (!me->arch.stubs_section) {
317 pr_err("%s: doesn't contain .stubs.\n", me->name);
318 return -ENOEXEC;
319 }
320
321 /* If we don't have a .toc, just use .stubs. We need to set r2
322 to some reasonable value in case the module calls out to
323 other functions via a stub, or if a function pointer escapes
324 the module by some means. */
325 if (!me->arch.toc_section)
326 me->arch.toc_section = me->arch.stubs_section;
327
328 /* Override the stubs size */
329 sechdrs[me->arch.stubs_section].sh_size = get_stubs_size(hdr, sechdrs);
330 return 0;
331}
332
333#ifdef CONFIG_MPROFILE_KERNEL
334
335static u32 stub_insns[] = {
336 PPC_RAW_LD(_R12, _R13, offsetof(struct paca_struct, kernel_toc)),
337 PPC_RAW_ADDIS(_R12, _R12, 0),
338 PPC_RAW_ADDI(_R12, _R12, 0),
339 PPC_RAW_MTCTR(_R12),
340 PPC_RAW_BCTR(),
341};
342
343/*
344 * For mprofile-kernel we use a special stub for ftrace_caller() because we
345 * can't rely on r2 containing this module's TOC when we enter the stub.
346 *
347 * That can happen if the function calling us didn't need to use the toc. In
348 * that case it won't have setup r2, and the r2 value will be either the
349 * kernel's toc, or possibly another modules toc.
350 *
351 * To deal with that this stub uses the kernel toc, which is always accessible
352 * via the paca (in r13). The target (ftrace_caller()) is responsible for
353 * saving and restoring the toc before returning.
354 */
355static inline int create_ftrace_stub(struct ppc64_stub_entry *entry,
356 unsigned long addr,
357 struct module *me)
358{
359 long reladdr;
360
361 memcpy(entry->jump, stub_insns, sizeof(stub_insns));
362
363 /* Stub uses address relative to kernel toc (from the paca) */
364 reladdr = addr - kernel_toc_addr();
365 if (reladdr > 0x7FFFFFFF || reladdr < -(0x80000000L)) {
366 pr_err("%s: Address of %ps out of range of kernel_toc.\n",
367 me->name, (void *)addr);
368 return 0;
369 }
370
371 entry->jump[1] |= PPC_HA(reladdr);
372 entry->jump[2] |= PPC_LO(reladdr);
373
374 /* Even though we don't use funcdata in the stub, it's needed elsewhere. */
375 entry->funcdata = func_desc(addr);
376 entry->magic = STUB_MAGIC;
377
378 return 1;
379}
380
381static bool is_mprofile_ftrace_call(const char *name)
382{
383 if (!strcmp("_mcount", name))
384 return true;
385#ifdef CONFIG_DYNAMIC_FTRACE
386 if (!strcmp("ftrace_caller", name))
387 return true;
388#ifdef CONFIG_DYNAMIC_FTRACE_WITH_REGS
389 if (!strcmp("ftrace_regs_caller", name))
390 return true;
391#endif
392#endif
393
394 return false;
395}
396#else
397static inline int create_ftrace_stub(struct ppc64_stub_entry *entry,
398 unsigned long addr,
399 struct module *me)
400{
401 return 0;
402}
403
404static bool is_mprofile_ftrace_call(const char *name)
405{
406 return false;
407}
408#endif
409
410/*
411 * r2 is the TOC pointer: it actually points 0x8000 into the TOC (this gives the
412 * value maximum span in an instruction which uses a signed offset). Round down
413 * to a 256 byte boundary for the odd case where we are setting up r2 without a
414 * .toc section.
415 */
416static inline unsigned long my_r2(const Elf64_Shdr *sechdrs, struct module *me)
417{
418 return (sechdrs[me->arch.toc_section].sh_addr & ~0xfful) + 0x8000;
419}
420
421/* Patch stub to reference function and correct r2 value. */
422static inline int create_stub(const Elf64_Shdr *sechdrs,
423 struct ppc64_stub_entry *entry,
424 unsigned long addr,
425 struct module *me,
426 const char *name)
427{
428 long reladdr;
429 func_desc_t desc;
430 int i;
431
432 if (is_mprofile_ftrace_call(name))
433 return create_ftrace_stub(entry, addr, me);
434
435 for (i = 0; i < ARRAY_SIZE(ppc64_stub_insns); i++) {
436 if (patch_instruction(&entry->jump[i],
437 ppc_inst(ppc64_stub_insns[i])))
438 return 0;
439 }
440
441 /* Stub uses address relative to r2. */
442 reladdr = (unsigned long)entry - my_r2(sechdrs, me);
443 if (reladdr > 0x7FFFFFFF || reladdr < -(0x80000000L)) {
444 pr_err("%s: Address %p of stub out of range of %p.\n",
445 me->name, (void *)reladdr, (void *)my_r2);
446 return 0;
447 }
448 pr_debug("Stub %p get data from reladdr %li\n", entry, reladdr);
449
450 if (patch_instruction(&entry->jump[0],
451 ppc_inst(entry->jump[0] | PPC_HA(reladdr))))
452 return 0;
453
454 if (patch_instruction(&entry->jump[1],
455 ppc_inst(entry->jump[1] | PPC_LO(reladdr))))
456 return 0;
457
458 // func_desc_t is 8 bytes if ABIv2, else 16 bytes
459 desc = func_desc(addr);
460 for (i = 0; i < sizeof(func_desc_t) / sizeof(u32); i++) {
461 if (patch_instruction(((u32 *)&entry->funcdata) + i,
462 ppc_inst(((u32 *)(&desc))[i])))
463 return 0;
464 }
465
466 if (patch_instruction(&entry->magic, ppc_inst(STUB_MAGIC)))
467 return 0;
468
469 return 1;
470}
471
472/* Create stub to jump to function described in this OPD/ptr: we need the
473 stub to set up the TOC ptr (r2) for the function. */
474static unsigned long stub_for_addr(const Elf64_Shdr *sechdrs,
475 unsigned long addr,
476 struct module *me,
477 const char *name)
478{
479 struct ppc64_stub_entry *stubs;
480 unsigned int i, num_stubs;
481
482 num_stubs = sechdrs[me->arch.stubs_section].sh_size / sizeof(*stubs);
483
484 /* Find this stub, or if that fails, the next avail. entry */
485 stubs = (void *)sechdrs[me->arch.stubs_section].sh_addr;
486 for (i = 0; stub_func_addr(stubs[i].funcdata); i++) {
487 if (WARN_ON(i >= num_stubs))
488 return 0;
489
490 if (stub_func_addr(stubs[i].funcdata) == func_addr(addr))
491 return (unsigned long)&stubs[i];
492 }
493
494 if (!create_stub(sechdrs, &stubs[i], addr, me, name))
495 return 0;
496
497 return (unsigned long)&stubs[i];
498}
499
500/* We expect a noop next: if it is, replace it with instruction to
501 restore r2. */
502static int restore_r2(const char *name, u32 *instruction, struct module *me)
503{
504 u32 *prev_insn = instruction - 1;
505
506 if (is_mprofile_ftrace_call(name))
507 return 1;
508
509 /*
510 * Make sure the branch isn't a sibling call. Sibling calls aren't
511 * "link" branches and they don't return, so they don't need the r2
512 * restore afterwards.
513 */
514 if (!instr_is_relative_link_branch(ppc_inst(*prev_insn)))
515 return 1;
516
517 if (*instruction != PPC_RAW_NOP()) {
518 pr_err("%s: Expected nop after call, got %08x at %pS\n",
519 me->name, *instruction, instruction);
520 return 0;
521 }
522
523 /* ld r2,R2_STACK_OFFSET(r1) */
524 if (patch_instruction(instruction, ppc_inst(PPC_INST_LD_TOC)))
525 return 0;
526
527 return 1;
528}
529
530int apply_relocate_add(Elf64_Shdr *sechdrs,
531 const char *strtab,
532 unsigned int symindex,
533 unsigned int relsec,
534 struct module *me)
535{
536 unsigned int i;
537 Elf64_Rela *rela = (void *)sechdrs[relsec].sh_addr;
538 Elf64_Sym *sym;
539 unsigned long *location;
540 unsigned long value;
541
542 pr_debug("Applying ADD relocate section %u to %u\n", relsec,
543 sechdrs[relsec].sh_info);
544
545 /* First time we're called, we can fix up .TOC. */
546 if (!me->arch.toc_fixed) {
547 sym = find_dot_toc(sechdrs, strtab, symindex);
548 /* It's theoretically possible that a module doesn't want a
549 * .TOC. so don't fail it just for that. */
550 if (sym)
551 sym->st_value = my_r2(sechdrs, me);
552 me->arch.toc_fixed = true;
553 }
554
555 for (i = 0; i < sechdrs[relsec].sh_size / sizeof(*rela); i++) {
556 /* This is where to make the change */
557 location = (void *)sechdrs[sechdrs[relsec].sh_info].sh_addr
558 + rela[i].r_offset;
559 /* This is the symbol it is referring to */
560 sym = (Elf64_Sym *)sechdrs[symindex].sh_addr
561 + ELF64_R_SYM(rela[i].r_info);
562
563 pr_debug("RELOC at %p: %li-type as %s (0x%lx) + %li\n",
564 location, (long)ELF64_R_TYPE(rela[i].r_info),
565 strtab + sym->st_name, (unsigned long)sym->st_value,
566 (long)rela[i].r_addend);
567
568 /* `Everything is relative'. */
569 value = sym->st_value + rela[i].r_addend;
570
571 switch (ELF64_R_TYPE(rela[i].r_info)) {
572 case R_PPC64_ADDR32:
573 /* Simply set it */
574 *(u32 *)location = value;
575 break;
576
577 case R_PPC64_ADDR64:
578 /* Simply set it */
579 *(unsigned long *)location = value;
580 break;
581
582 case R_PPC64_TOC:
583 *(unsigned long *)location = my_r2(sechdrs, me);
584 break;
585
586 case R_PPC64_TOC16:
587 /* Subtract TOC pointer */
588 value -= my_r2(sechdrs, me);
589 if (value + 0x8000 > 0xffff) {
590 pr_err("%s: bad TOC16 relocation (0x%lx)\n",
591 me->name, value);
592 return -ENOEXEC;
593 }
594 *((uint16_t *) location)
595 = (*((uint16_t *) location) & ~0xffff)
596 | (value & 0xffff);
597 break;
598
599 case R_PPC64_TOC16_LO:
600 /* Subtract TOC pointer */
601 value -= my_r2(sechdrs, me);
602 *((uint16_t *) location)
603 = (*((uint16_t *) location) & ~0xffff)
604 | (value & 0xffff);
605 break;
606
607 case R_PPC64_TOC16_DS:
608 /* Subtract TOC pointer */
609 value -= my_r2(sechdrs, me);
610 if ((value & 3) != 0 || value + 0x8000 > 0xffff) {
611 pr_err("%s: bad TOC16_DS relocation (0x%lx)\n",
612 me->name, value);
613 return -ENOEXEC;
614 }
615 *((uint16_t *) location)
616 = (*((uint16_t *) location) & ~0xfffc)
617 | (value & 0xfffc);
618 break;
619
620 case R_PPC64_TOC16_LO_DS:
621 /* Subtract TOC pointer */
622 value -= my_r2(sechdrs, me);
623 if ((value & 3) != 0) {
624 pr_err("%s: bad TOC16_LO_DS relocation (0x%lx)\n",
625 me->name, value);
626 return -ENOEXEC;
627 }
628 *((uint16_t *) location)
629 = (*((uint16_t *) location) & ~0xfffc)
630 | (value & 0xfffc);
631 break;
632
633 case R_PPC64_TOC16_HA:
634 /* Subtract TOC pointer */
635 value -= my_r2(sechdrs, me);
636 value = ((value + 0x8000) >> 16);
637 *((uint16_t *) location)
638 = (*((uint16_t *) location) & ~0xffff)
639 | (value & 0xffff);
640 break;
641
642 case R_PPC_REL24:
643 /* FIXME: Handle weak symbols here --RR */
644 if (sym->st_shndx == SHN_UNDEF ||
645 sym->st_shndx == SHN_LIVEPATCH) {
646 /* External: go via stub */
647 value = stub_for_addr(sechdrs, value, me,
648 strtab + sym->st_name);
649 if (!value)
650 return -ENOENT;
651 if (!restore_r2(strtab + sym->st_name,
652 (u32 *)location + 1, me))
653 return -ENOEXEC;
654 } else
655 value += local_entry_offset(sym);
656
657 /* Convert value to relative */
658 value -= (unsigned long)location;
659 if (value + 0x2000000 > 0x3ffffff || (value & 3) != 0){
660 pr_err("%s: REL24 %li out of range!\n",
661 me->name, (long int)value);
662 return -ENOEXEC;
663 }
664
665 /* Only replace bits 2 through 26 */
666 value = (*(uint32_t *)location & ~PPC_LI_MASK) | PPC_LI(value);
667
668 if (patch_instruction((u32 *)location, ppc_inst(value)))
669 return -EFAULT;
670
671 break;
672
673 case R_PPC64_REL64:
674 /* 64 bits relative (used by features fixups) */
675 *location = value - (unsigned long)location;
676 break;
677
678 case R_PPC64_REL32:
679 /* 32 bits relative (used by relative exception tables) */
680 /* Convert value to relative */
681 value -= (unsigned long)location;
682 if (value + 0x80000000 > 0xffffffff) {
683 pr_err("%s: REL32 %li out of range!\n",
684 me->name, (long int)value);
685 return -ENOEXEC;
686 }
687 *(u32 *)location = value;
688 break;
689
690 case R_PPC64_TOCSAVE:
691 /*
692 * Marker reloc indicates we don't have to save r2.
693 * That would only save us one instruction, so ignore
694 * it.
695 */
696 break;
697
698 case R_PPC64_ENTRY:
699 /*
700 * Optimize ELFv2 large code model entry point if
701 * the TOC is within 2GB range of current location.
702 */
703 value = my_r2(sechdrs, me) - (unsigned long)location;
704 if (value + 0x80008000 > 0xffffffff)
705 break;
706 /*
707 * Check for the large code model prolog sequence:
708 * ld r2, ...(r12)
709 * add r2, r2, r12
710 */
711 if ((((uint32_t *)location)[0] & ~0xfffc) != PPC_RAW_LD(_R2, _R12, 0))
712 break;
713 if (((uint32_t *)location)[1] != PPC_RAW_ADD(_R2, _R2, _R12))
714 break;
715 /*
716 * If found, replace it with:
717 * addis r2, r12, (.TOC.-func)@ha
718 * addi r2, r2, (.TOC.-func)@l
719 */
720 ((uint32_t *)location)[0] = PPC_RAW_ADDIS(_R2, _R12, PPC_HA(value));
721 ((uint32_t *)location)[1] = PPC_RAW_ADDI(_R2, _R2, PPC_LO(value));
722 break;
723
724 case R_PPC64_REL16_HA:
725 /* Subtract location pointer */
726 value -= (unsigned long)location;
727 value = ((value + 0x8000) >> 16);
728 *((uint16_t *) location)
729 = (*((uint16_t *) location) & ~0xffff)
730 | (value & 0xffff);
731 break;
732
733 case R_PPC64_REL16_LO:
734 /* Subtract location pointer */
735 value -= (unsigned long)location;
736 *((uint16_t *) location)
737 = (*((uint16_t *) location) & ~0xffff)
738 | (value & 0xffff);
739 break;
740
741 default:
742 pr_err("%s: Unknown ADD relocation: %lu\n",
743 me->name,
744 (unsigned long)ELF64_R_TYPE(rela[i].r_info));
745 return -ENOEXEC;
746 }
747 }
748
749 return 0;
750}
751
752#ifdef CONFIG_DYNAMIC_FTRACE
753int module_trampoline_target(struct module *mod, unsigned long addr,
754 unsigned long *target)
755{
756 struct ppc64_stub_entry *stub;
757 func_desc_t funcdata;
758 u32 magic;
759
760 if (!within_module_core(addr, mod)) {
761 pr_err("%s: stub %lx not in module %s\n", __func__, addr, mod->name);
762 return -EFAULT;
763 }
764
765 stub = (struct ppc64_stub_entry *)addr;
766
767 if (copy_from_kernel_nofault(&magic, &stub->magic,
768 sizeof(magic))) {
769 pr_err("%s: fault reading magic for stub %lx for %s\n", __func__, addr, mod->name);
770 return -EFAULT;
771 }
772
773 if (magic != STUB_MAGIC) {
774 pr_err("%s: bad magic for stub %lx for %s\n", __func__, addr, mod->name);
775 return -EFAULT;
776 }
777
778 if (copy_from_kernel_nofault(&funcdata, &stub->funcdata,
779 sizeof(funcdata))) {
780 pr_err("%s: fault reading funcdata for stub %lx for %s\n", __func__, addr, mod->name);
781 return -EFAULT;
782 }
783
784 *target = stub_func_addr(funcdata);
785
786 return 0;
787}
788
789int module_finalize_ftrace(struct module *mod, const Elf_Shdr *sechdrs)
790{
791 mod->arch.tramp = stub_for_addr(sechdrs,
792 (unsigned long)ftrace_caller,
793 mod,
794 "ftrace_caller");
795#ifdef CONFIG_DYNAMIC_FTRACE_WITH_REGS
796 mod->arch.tramp_regs = stub_for_addr(sechdrs,
797 (unsigned long)ftrace_regs_caller,
798 mod,
799 "ftrace_regs_caller");
800 if (!mod->arch.tramp_regs)
801 return -ENOENT;
802#endif
803
804 if (!mod->arch.tramp)
805 return -ENOENT;
806
807 return 0;
808}
809#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 <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
24/* FIXME: We don't do .init separately. To do this, we'd need to have
25 a separate r2 value in the init and core section, and stub between
26 them, too.
27
28 Using a magic allocator which places modules within 32MB solves
29 this, and makes other things simpler. Anton?
30 --RR. */
31
32#ifdef PPC64_ELF_ABI_v2
33
34/* An address is simply the address of the function. */
35typedef unsigned long func_desc_t;
36
37static func_desc_t func_desc(unsigned long addr)
38{
39 return addr;
40}
41static unsigned long func_addr(unsigned long addr)
42{
43 return addr;
44}
45static unsigned long stub_func_addr(func_desc_t func)
46{
47 return func;
48}
49
50/* PowerPC64 specific values for the Elf64_Sym st_other field. */
51#define STO_PPC64_LOCAL_BIT 5
52#define STO_PPC64_LOCAL_MASK (7 << STO_PPC64_LOCAL_BIT)
53#define PPC64_LOCAL_ENTRY_OFFSET(other) \
54 (((1 << (((other) & STO_PPC64_LOCAL_MASK) >> STO_PPC64_LOCAL_BIT)) >> 2) << 2)
55
56static unsigned int local_entry_offset(const Elf64_Sym *sym)
57{
58 /* sym->st_other indicates offset to local entry point
59 * (otherwise it will assume r12 is the address of the start
60 * of function and try to derive r2 from it). */
61 return PPC64_LOCAL_ENTRY_OFFSET(sym->st_other);
62}
63#else
64
65/* An address is address of the OPD entry, which contains address of fn. */
66typedef struct ppc64_opd_entry func_desc_t;
67
68static func_desc_t func_desc(unsigned long addr)
69{
70 return *(struct ppc64_opd_entry *)addr;
71}
72static unsigned long func_addr(unsigned long addr)
73{
74 return func_desc(addr).funcaddr;
75}
76static unsigned long stub_func_addr(func_desc_t func)
77{
78 return func.funcaddr;
79}
80static unsigned int local_entry_offset(const Elf64_Sym *sym)
81{
82 return 0;
83}
84
85void *dereference_module_function_descriptor(struct module *mod, void *ptr)
86{
87 if (ptr < (void *)mod->arch.start_opd ||
88 ptr >= (void *)mod->arch.end_opd)
89 return ptr;
90
91 return dereference_function_descriptor(ptr);
92}
93#endif
94
95#define STUB_MAGIC 0x73747562 /* stub */
96
97/* Like PPC32, we need little trampolines to do > 24-bit jumps (into
98 the kernel itself). But on PPC64, these need to be used for every
99 jump, actually, to reset r2 (TOC+0x8000). */
100struct ppc64_stub_entry
101{
102 /* 28 byte jump instruction sequence (7 instructions). We only
103 * need 6 instructions on ABIv2 but we always allocate 7 so
104 * so we don't have to modify the trampoline load instruction. */
105 u32 jump[7];
106 /* Used by ftrace to identify stubs */
107 u32 magic;
108 /* Data for the above code */
109 func_desc_t funcdata;
110};
111
112/*
113 * PPC64 uses 24 bit jumps, but we need to jump into other modules or
114 * the kernel which may be further. So we jump to a stub.
115 *
116 * For ELFv1 we need to use this to set up the new r2 value (aka TOC
117 * pointer). For ELFv2 it's the callee's responsibility to set up the
118 * new r2, but for both we need to save the old r2.
119 *
120 * We could simply patch the new r2 value and function pointer into
121 * the stub, but it's significantly shorter to put these values at the
122 * end of the stub code, and patch the stub address (32-bits relative
123 * to the TOC ptr, r2) into the stub.
124 *
125 * addis r11,r2, <high>
126 * addi r11,r11, <low>
127 * std r2,R2_STACK_OFFSET(r1)
128 * ld r12,32(r11)
129 * ld r2,40(r11)
130 * mtctr r12
131 * bctr
132 */
133static u32 ppc64_stub_insns[] = {
134 PPC_INST_ADDIS | __PPC_RT(R11) | __PPC_RA(R2),
135 PPC_INST_ADDI | __PPC_RT(R11) | __PPC_RA(R11),
136 /* Save current r2 value in magic place on the stack. */
137 PPC_INST_STD | __PPC_RS(R2) | __PPC_RA(R1) | R2_STACK_OFFSET,
138 PPC_INST_LD | __PPC_RT(R12) | __PPC_RA(R11) | 32,
139#ifdef PPC64_ELF_ABI_v1
140 /* Set up new r2 from function descriptor */
141 PPC_INST_LD | __PPC_RT(R2) | __PPC_RA(R11) | 40,
142#endif
143 PPC_INST_MTCTR | __PPC_RS(R12),
144 PPC_INST_BCTR,
145};
146
147#ifdef CONFIG_DYNAMIC_FTRACE
148int module_trampoline_target(struct module *mod, unsigned long addr,
149 unsigned long *target)
150{
151 struct ppc64_stub_entry *stub;
152 func_desc_t funcdata;
153 u32 magic;
154
155 if (!within_module_core(addr, mod)) {
156 pr_err("%s: stub %lx not in module %s\n", __func__, addr, mod->name);
157 return -EFAULT;
158 }
159
160 stub = (struct ppc64_stub_entry *)addr;
161
162 if (probe_kernel_read(&magic, &stub->magic, sizeof(magic))) {
163 pr_err("%s: fault reading magic for stub %lx for %s\n", __func__, addr, mod->name);
164 return -EFAULT;
165 }
166
167 if (magic != STUB_MAGIC) {
168 pr_err("%s: bad magic for stub %lx for %s\n", __func__, addr, mod->name);
169 return -EFAULT;
170 }
171
172 if (probe_kernel_read(&funcdata, &stub->funcdata, sizeof(funcdata))) {
173 pr_err("%s: fault reading funcdata for stub %lx for %s\n", __func__, addr, mod->name);
174 return -EFAULT;
175 }
176
177 *target = stub_func_addr(funcdata);
178
179 return 0;
180}
181#endif
182
183/* Count how many different 24-bit relocations (different symbol,
184 different addend) */
185static unsigned int count_relocs(const Elf64_Rela *rela, unsigned int num)
186{
187 unsigned int i, r_info, r_addend, _count_relocs;
188
189 /* FIXME: Only count external ones --RR */
190 _count_relocs = 0;
191 r_info = 0;
192 r_addend = 0;
193 for (i = 0; i < num; i++)
194 /* Only count 24-bit relocs, others don't need stubs */
195 if (ELF64_R_TYPE(rela[i].r_info) == R_PPC_REL24 &&
196 (r_info != ELF64_R_SYM(rela[i].r_info) ||
197 r_addend != rela[i].r_addend)) {
198 _count_relocs++;
199 r_info = ELF64_R_SYM(rela[i].r_info);
200 r_addend = rela[i].r_addend;
201 }
202
203 return _count_relocs;
204}
205
206static int relacmp(const void *_x, const void *_y)
207{
208 const Elf64_Rela *x, *y;
209
210 y = (Elf64_Rela *)_x;
211 x = (Elf64_Rela *)_y;
212
213 /* Compare the entire r_info (as opposed to ELF64_R_SYM(r_info) only) to
214 * make the comparison cheaper/faster. It won't affect the sorting or
215 * the counting algorithms' performance
216 */
217 if (x->r_info < y->r_info)
218 return -1;
219 else if (x->r_info > y->r_info)
220 return 1;
221 else if (x->r_addend < y->r_addend)
222 return -1;
223 else if (x->r_addend > y->r_addend)
224 return 1;
225 else
226 return 0;
227}
228
229static void relaswap(void *_x, void *_y, int size)
230{
231 uint64_t *x, *y, tmp;
232 int i;
233
234 y = (uint64_t *)_x;
235 x = (uint64_t *)_y;
236
237 for (i = 0; i < sizeof(Elf64_Rela) / sizeof(uint64_t); i++) {
238 tmp = x[i];
239 x[i] = y[i];
240 y[i] = tmp;
241 }
242}
243
244/* Get size of potential trampolines required. */
245static unsigned long get_stubs_size(const Elf64_Ehdr *hdr,
246 const Elf64_Shdr *sechdrs)
247{
248 /* One extra reloc so it's always 0-funcaddr terminated */
249 unsigned long relocs = 1;
250 unsigned i;
251
252 /* Every relocated section... */
253 for (i = 1; i < hdr->e_shnum; i++) {
254 if (sechdrs[i].sh_type == SHT_RELA) {
255 pr_debug("Found relocations in section %u\n", i);
256 pr_debug("Ptr: %p. Number: %Lu\n",
257 (void *)sechdrs[i].sh_addr,
258 sechdrs[i].sh_size / sizeof(Elf64_Rela));
259
260 /* Sort the relocation information based on a symbol and
261 * addend key. This is a stable O(n*log n) complexity
262 * alogrithm but it will reduce the complexity of
263 * count_relocs() to linear complexity O(n)
264 */
265 sort((void *)sechdrs[i].sh_addr,
266 sechdrs[i].sh_size / sizeof(Elf64_Rela),
267 sizeof(Elf64_Rela), relacmp, relaswap);
268
269 relocs += count_relocs((void *)sechdrs[i].sh_addr,
270 sechdrs[i].sh_size
271 / sizeof(Elf64_Rela));
272 }
273 }
274
275#ifdef CONFIG_DYNAMIC_FTRACE
276 /* make the trampoline to the ftrace_caller */
277 relocs++;
278#ifdef CONFIG_DYNAMIC_FTRACE_WITH_REGS
279 /* an additional one for ftrace_regs_caller */
280 relocs++;
281#endif
282#endif
283
284 pr_debug("Looks like a total of %lu stubs, max\n", relocs);
285 return relocs * sizeof(struct ppc64_stub_entry);
286}
287
288/* Still needed for ELFv2, for .TOC. */
289static void dedotify_versions(struct modversion_info *vers,
290 unsigned long size)
291{
292 struct modversion_info *end;
293
294 for (end = (void *)vers + size; vers < end; vers++)
295 if (vers->name[0] == '.') {
296 memmove(vers->name, vers->name+1, strlen(vers->name));
297 }
298}
299
300/*
301 * Undefined symbols which refer to .funcname, hack to funcname. Make .TOC.
302 * seem to be defined (value set later).
303 */
304static void dedotify(Elf64_Sym *syms, unsigned int numsyms, char *strtab)
305{
306 unsigned int i;
307
308 for (i = 1; i < numsyms; i++) {
309 if (syms[i].st_shndx == SHN_UNDEF) {
310 char *name = strtab + syms[i].st_name;
311 if (name[0] == '.') {
312 if (strcmp(name+1, "TOC.") == 0)
313 syms[i].st_shndx = SHN_ABS;
314 syms[i].st_name++;
315 }
316 }
317 }
318}
319
320static Elf64_Sym *find_dot_toc(Elf64_Shdr *sechdrs,
321 const char *strtab,
322 unsigned int symindex)
323{
324 unsigned int i, numsyms;
325 Elf64_Sym *syms;
326
327 syms = (Elf64_Sym *)sechdrs[symindex].sh_addr;
328 numsyms = sechdrs[symindex].sh_size / sizeof(Elf64_Sym);
329
330 for (i = 1; i < numsyms; i++) {
331 if (syms[i].st_shndx == SHN_ABS
332 && strcmp(strtab + syms[i].st_name, "TOC.") == 0)
333 return &syms[i];
334 }
335 return NULL;
336}
337
338int module_frob_arch_sections(Elf64_Ehdr *hdr,
339 Elf64_Shdr *sechdrs,
340 char *secstrings,
341 struct module *me)
342{
343 unsigned int i;
344
345 /* Find .toc and .stubs sections, symtab and strtab */
346 for (i = 1; i < hdr->e_shnum; i++) {
347 char *p;
348 if (strcmp(secstrings + sechdrs[i].sh_name, ".stubs") == 0)
349 me->arch.stubs_section = i;
350 else if (strcmp(secstrings + sechdrs[i].sh_name, ".toc") == 0) {
351 me->arch.toc_section = i;
352 if (sechdrs[i].sh_addralign < 8)
353 sechdrs[i].sh_addralign = 8;
354 }
355 else if (strcmp(secstrings+sechdrs[i].sh_name,"__versions")==0)
356 dedotify_versions((void *)hdr + sechdrs[i].sh_offset,
357 sechdrs[i].sh_size);
358
359 /* We don't handle .init for the moment: rename to _init */
360 while ((p = strstr(secstrings + sechdrs[i].sh_name, ".init")))
361 p[0] = '_';
362
363 if (sechdrs[i].sh_type == SHT_SYMTAB)
364 dedotify((void *)hdr + sechdrs[i].sh_offset,
365 sechdrs[i].sh_size / sizeof(Elf64_Sym),
366 (void *)hdr
367 + sechdrs[sechdrs[i].sh_link].sh_offset);
368 }
369
370 if (!me->arch.stubs_section) {
371 pr_err("%s: doesn't contain .stubs.\n", me->name);
372 return -ENOEXEC;
373 }
374
375 /* If we don't have a .toc, just use .stubs. We need to set r2
376 to some reasonable value in case the module calls out to
377 other functions via a stub, or if a function pointer escapes
378 the module by some means. */
379 if (!me->arch.toc_section)
380 me->arch.toc_section = me->arch.stubs_section;
381
382 /* Override the stubs size */
383 sechdrs[me->arch.stubs_section].sh_size = get_stubs_size(hdr, sechdrs);
384 return 0;
385}
386
387/*
388 * r2 is the TOC pointer: it actually points 0x8000 into the TOC (this gives the
389 * value maximum span in an instruction which uses a signed offset). Round down
390 * to a 256 byte boundary for the odd case where we are setting up r2 without a
391 * .toc section.
392 */
393static inline unsigned long my_r2(const Elf64_Shdr *sechdrs, struct module *me)
394{
395 return (sechdrs[me->arch.toc_section].sh_addr & ~0xfful) + 0x8000;
396}
397
398/* Patch stub to reference function and correct r2 value. */
399static inline int create_stub(const Elf64_Shdr *sechdrs,
400 struct ppc64_stub_entry *entry,
401 unsigned long addr,
402 struct module *me)
403{
404 long reladdr;
405
406 memcpy(entry->jump, ppc64_stub_insns, sizeof(ppc64_stub_insns));
407
408 /* Stub uses address relative to r2. */
409 reladdr = (unsigned long)entry - my_r2(sechdrs, me);
410 if (reladdr > 0x7FFFFFFF || reladdr < -(0x80000000L)) {
411 pr_err("%s: Address %p of stub out of range of %p.\n",
412 me->name, (void *)reladdr, (void *)my_r2);
413 return 0;
414 }
415 pr_debug("Stub %p get data from reladdr %li\n", entry, reladdr);
416
417 entry->jump[0] |= PPC_HA(reladdr);
418 entry->jump[1] |= PPC_LO(reladdr);
419 entry->funcdata = func_desc(addr);
420 entry->magic = STUB_MAGIC;
421
422 return 1;
423}
424
425/* Create stub to jump to function described in this OPD/ptr: we need the
426 stub to set up the TOC ptr (r2) for the function. */
427static unsigned long stub_for_addr(const Elf64_Shdr *sechdrs,
428 unsigned long addr,
429 struct module *me)
430{
431 struct ppc64_stub_entry *stubs;
432 unsigned int i, num_stubs;
433
434 num_stubs = sechdrs[me->arch.stubs_section].sh_size / sizeof(*stubs);
435
436 /* Find this stub, or if that fails, the next avail. entry */
437 stubs = (void *)sechdrs[me->arch.stubs_section].sh_addr;
438 for (i = 0; stub_func_addr(stubs[i].funcdata); i++) {
439 if (WARN_ON(i >= num_stubs))
440 return 0;
441
442 if (stub_func_addr(stubs[i].funcdata) == func_addr(addr))
443 return (unsigned long)&stubs[i];
444 }
445
446 if (!create_stub(sechdrs, &stubs[i], addr, me))
447 return 0;
448
449 return (unsigned long)&stubs[i];
450}
451
452#ifdef CONFIG_MPROFILE_KERNEL
453static bool is_mprofile_mcount_callsite(const char *name, u32 *instruction)
454{
455 if (strcmp("_mcount", name))
456 return false;
457
458 /*
459 * Check if this is one of the -mprofile-kernel sequences.
460 */
461 if (instruction[-1] == PPC_INST_STD_LR &&
462 instruction[-2] == PPC_INST_MFLR)
463 return true;
464
465 if (instruction[-1] == PPC_INST_MFLR)
466 return true;
467
468 return false;
469}
470
471/*
472 * In case of _mcount calls, do not save the current callee's TOC (in r2) into
473 * the original caller's stack frame. If we did we would clobber the saved TOC
474 * value of the original caller.
475 */
476static void squash_toc_save_inst(const char *name, unsigned long addr)
477{
478 struct ppc64_stub_entry *stub = (struct ppc64_stub_entry *)addr;
479
480 /* Only for calls to _mcount */
481 if (strcmp("_mcount", name) != 0)
482 return;
483
484 stub->jump[2] = PPC_INST_NOP;
485}
486#else
487static void squash_toc_save_inst(const char *name, unsigned long addr) { }
488
489static bool is_mprofile_mcount_callsite(const char *name, u32 *instruction)
490{
491 return false;
492}
493#endif
494
495/* We expect a noop next: if it is, replace it with instruction to
496 restore r2. */
497static int restore_r2(const char *name, u32 *instruction, struct module *me)
498{
499 u32 *prev_insn = instruction - 1;
500
501 if (is_mprofile_mcount_callsite(name, prev_insn))
502 return 1;
503
504 /*
505 * Make sure the branch isn't a sibling call. Sibling calls aren't
506 * "link" branches and they don't return, so they don't need the r2
507 * restore afterwards.
508 */
509 if (!instr_is_relative_link_branch(*prev_insn))
510 return 1;
511
512 if (*instruction != PPC_INST_NOP) {
513 pr_err("%s: Expected nop after call, got %08x at %pS\n",
514 me->name, *instruction, instruction);
515 return 0;
516 }
517 /* ld r2,R2_STACK_OFFSET(r1) */
518 *instruction = PPC_INST_LD_TOC;
519 return 1;
520}
521
522int apply_relocate_add(Elf64_Shdr *sechdrs,
523 const char *strtab,
524 unsigned int symindex,
525 unsigned int relsec,
526 struct module *me)
527{
528 unsigned int i;
529 Elf64_Rela *rela = (void *)sechdrs[relsec].sh_addr;
530 Elf64_Sym *sym;
531 unsigned long *location;
532 unsigned long value;
533
534 pr_debug("Applying ADD relocate section %u to %u\n", relsec,
535 sechdrs[relsec].sh_info);
536
537 /* First time we're called, we can fix up .TOC. */
538 if (!me->arch.toc_fixed) {
539 sym = find_dot_toc(sechdrs, strtab, symindex);
540 /* It's theoretically possible that a module doesn't want a
541 * .TOC. so don't fail it just for that. */
542 if (sym)
543 sym->st_value = my_r2(sechdrs, me);
544 me->arch.toc_fixed = true;
545 }
546
547 for (i = 0; i < sechdrs[relsec].sh_size / sizeof(*rela); i++) {
548 /* This is where to make the change */
549 location = (void *)sechdrs[sechdrs[relsec].sh_info].sh_addr
550 + rela[i].r_offset;
551 /* This is the symbol it is referring to */
552 sym = (Elf64_Sym *)sechdrs[symindex].sh_addr
553 + ELF64_R_SYM(rela[i].r_info);
554
555 pr_debug("RELOC at %p: %li-type as %s (0x%lx) + %li\n",
556 location, (long)ELF64_R_TYPE(rela[i].r_info),
557 strtab + sym->st_name, (unsigned long)sym->st_value,
558 (long)rela[i].r_addend);
559
560 /* `Everything is relative'. */
561 value = sym->st_value + rela[i].r_addend;
562
563 switch (ELF64_R_TYPE(rela[i].r_info)) {
564 case R_PPC64_ADDR32:
565 /* Simply set it */
566 *(u32 *)location = value;
567 break;
568
569 case R_PPC64_ADDR64:
570 /* Simply set it */
571 *(unsigned long *)location = value;
572 break;
573
574 case R_PPC64_TOC:
575 *(unsigned long *)location = my_r2(sechdrs, me);
576 break;
577
578 case R_PPC64_TOC16:
579 /* Subtract TOC pointer */
580 value -= my_r2(sechdrs, me);
581 if (value + 0x8000 > 0xffff) {
582 pr_err("%s: bad TOC16 relocation (0x%lx)\n",
583 me->name, value);
584 return -ENOEXEC;
585 }
586 *((uint16_t *) location)
587 = (*((uint16_t *) location) & ~0xffff)
588 | (value & 0xffff);
589 break;
590
591 case R_PPC64_TOC16_LO:
592 /* Subtract TOC pointer */
593 value -= my_r2(sechdrs, me);
594 *((uint16_t *) location)
595 = (*((uint16_t *) location) & ~0xffff)
596 | (value & 0xffff);
597 break;
598
599 case R_PPC64_TOC16_DS:
600 /* Subtract TOC pointer */
601 value -= my_r2(sechdrs, me);
602 if ((value & 3) != 0 || value + 0x8000 > 0xffff) {
603 pr_err("%s: bad TOC16_DS relocation (0x%lx)\n",
604 me->name, value);
605 return -ENOEXEC;
606 }
607 *((uint16_t *) location)
608 = (*((uint16_t *) location) & ~0xfffc)
609 | (value & 0xfffc);
610 break;
611
612 case R_PPC64_TOC16_LO_DS:
613 /* Subtract TOC pointer */
614 value -= my_r2(sechdrs, me);
615 if ((value & 3) != 0) {
616 pr_err("%s: bad TOC16_LO_DS relocation (0x%lx)\n",
617 me->name, value);
618 return -ENOEXEC;
619 }
620 *((uint16_t *) location)
621 = (*((uint16_t *) location) & ~0xfffc)
622 | (value & 0xfffc);
623 break;
624
625 case R_PPC64_TOC16_HA:
626 /* Subtract TOC pointer */
627 value -= my_r2(sechdrs, me);
628 value = ((value + 0x8000) >> 16);
629 *((uint16_t *) location)
630 = (*((uint16_t *) location) & ~0xffff)
631 | (value & 0xffff);
632 break;
633
634 case R_PPC_REL24:
635 /* FIXME: Handle weak symbols here --RR */
636 if (sym->st_shndx == SHN_UNDEF ||
637 sym->st_shndx == SHN_LIVEPATCH) {
638 /* External: go via stub */
639 value = stub_for_addr(sechdrs, value, me);
640 if (!value)
641 return -ENOENT;
642 if (!restore_r2(strtab + sym->st_name,
643 (u32 *)location + 1, me))
644 return -ENOEXEC;
645
646 squash_toc_save_inst(strtab + sym->st_name, value);
647 } else
648 value += local_entry_offset(sym);
649
650 /* Convert value to relative */
651 value -= (unsigned long)location;
652 if (value + 0x2000000 > 0x3ffffff || (value & 3) != 0){
653 pr_err("%s: REL24 %li out of range!\n",
654 me->name, (long int)value);
655 return -ENOEXEC;
656 }
657
658 /* Only replace bits 2 through 26 */
659 *(uint32_t *)location
660 = (*(uint32_t *)location & ~0x03fffffc)
661 | (value & 0x03fffffc);
662 break;
663
664 case R_PPC64_REL64:
665 /* 64 bits relative (used by features fixups) */
666 *location = value - (unsigned long)location;
667 break;
668
669 case R_PPC64_REL32:
670 /* 32 bits relative (used by relative exception tables) */
671 /* Convert value to relative */
672 value -= (unsigned long)location;
673 if (value + 0x80000000 > 0xffffffff) {
674 pr_err("%s: REL32 %li out of range!\n",
675 me->name, (long int)value);
676 return -ENOEXEC;
677 }
678 *(u32 *)location = value;
679 break;
680
681 case R_PPC64_TOCSAVE:
682 /*
683 * Marker reloc indicates we don't have to save r2.
684 * That would only save us one instruction, so ignore
685 * it.
686 */
687 break;
688
689 case R_PPC64_ENTRY:
690 /*
691 * Optimize ELFv2 large code model entry point if
692 * the TOC is within 2GB range of current location.
693 */
694 value = my_r2(sechdrs, me) - (unsigned long)location;
695 if (value + 0x80008000 > 0xffffffff)
696 break;
697 /*
698 * Check for the large code model prolog sequence:
699 * ld r2, ...(r12)
700 * add r2, r2, r12
701 */
702 if ((((uint32_t *)location)[0] & ~0xfffc) !=
703 (PPC_INST_LD | __PPC_RT(R2) | __PPC_RA(R12)))
704 break;
705 if (((uint32_t *)location)[1] !=
706 (PPC_INST_ADD | __PPC_RT(R2) | __PPC_RA(R2) | __PPC_RB(R12)))
707 break;
708 /*
709 * If found, replace it with:
710 * addis r2, r12, (.TOC.-func)@ha
711 * addi r2, r2, (.TOC.-func)@l
712 */
713 ((uint32_t *)location)[0] = PPC_INST_ADDIS | __PPC_RT(R2) |
714 __PPC_RA(R12) | PPC_HA(value);
715 ((uint32_t *)location)[1] = PPC_INST_ADDI | __PPC_RT(R2) |
716 __PPC_RA(R2) | PPC_LO(value);
717 break;
718
719 case R_PPC64_REL16_HA:
720 /* Subtract location pointer */
721 value -= (unsigned long)location;
722 value = ((value + 0x8000) >> 16);
723 *((uint16_t *) location)
724 = (*((uint16_t *) location) & ~0xffff)
725 | (value & 0xffff);
726 break;
727
728 case R_PPC64_REL16_LO:
729 /* Subtract location pointer */
730 value -= (unsigned long)location;
731 *((uint16_t *) location)
732 = (*((uint16_t *) location) & ~0xffff)
733 | (value & 0xffff);
734 break;
735
736 default:
737 pr_err("%s: Unknown ADD relocation: %lu\n",
738 me->name,
739 (unsigned long)ELF64_R_TYPE(rela[i].r_info));
740 return -ENOEXEC;
741 }
742 }
743
744 return 0;
745}
746
747#ifdef CONFIG_DYNAMIC_FTRACE
748
749#ifdef CONFIG_MPROFILE_KERNEL
750
751#define PACATOC offsetof(struct paca_struct, kernel_toc)
752
753/*
754 * For mprofile-kernel we use a special stub for ftrace_caller() because we
755 * can't rely on r2 containing this module's TOC when we enter the stub.
756 *
757 * That can happen if the function calling us didn't need to use the toc. In
758 * that case it won't have setup r2, and the r2 value will be either the
759 * kernel's toc, or possibly another modules toc.
760 *
761 * To deal with that this stub uses the kernel toc, which is always accessible
762 * via the paca (in r13). The target (ftrace_caller()) is responsible for
763 * saving and restoring the toc before returning.
764 */
765static unsigned long create_ftrace_stub(const Elf64_Shdr *sechdrs,
766 struct module *me, unsigned long addr)
767{
768 struct ppc64_stub_entry *entry;
769 unsigned int i, num_stubs;
770 /*
771 * ld r12,PACATOC(r13)
772 * addis r12,r12,<high>
773 * addi r12,r12,<low>
774 * mtctr r12
775 * bctr
776 */
777 static u32 stub_insns[] = {
778 PPC_INST_LD | __PPC_RT(R12) | __PPC_RA(R13) | PACATOC,
779 PPC_INST_ADDIS | __PPC_RT(R12) | __PPC_RA(R12),
780 PPC_INST_ADDI | __PPC_RT(R12) | __PPC_RA(R12),
781 PPC_INST_MTCTR | __PPC_RS(R12),
782 PPC_INST_BCTR,
783 };
784 long reladdr;
785
786 num_stubs = sechdrs[me->arch.stubs_section].sh_size / sizeof(*entry);
787
788 /* Find the next available stub entry */
789 entry = (void *)sechdrs[me->arch.stubs_section].sh_addr;
790 for (i = 0; i < num_stubs && stub_func_addr(entry->funcdata); i++, entry++);
791
792 if (i >= num_stubs) {
793 pr_err("%s: Unable to find a free slot for ftrace stub.\n", me->name);
794 return 0;
795 }
796
797 memcpy(entry->jump, stub_insns, sizeof(stub_insns));
798
799 /* Stub uses address relative to kernel toc (from the paca) */
800 reladdr = addr - kernel_toc_addr();
801 if (reladdr > 0x7FFFFFFF || reladdr < -(0x80000000L)) {
802 pr_err("%s: Address of %ps out of range of kernel_toc.\n",
803 me->name, (void *)addr);
804 return 0;
805 }
806
807 entry->jump[1] |= PPC_HA(reladdr);
808 entry->jump[2] |= PPC_LO(reladdr);
809
810 /* Eventhough we don't use funcdata in the stub, it's needed elsewhere. */
811 entry->funcdata = func_desc(addr);
812 entry->magic = STUB_MAGIC;
813
814 return (unsigned long)entry;
815}
816#else
817static unsigned long create_ftrace_stub(const Elf64_Shdr *sechdrs,
818 struct module *me, unsigned long addr)
819{
820 return stub_for_addr(sechdrs, addr, me);
821}
822#endif
823
824int module_finalize_ftrace(struct module *mod, const Elf_Shdr *sechdrs)
825{
826 mod->arch.tramp = create_ftrace_stub(sechdrs, mod,
827 (unsigned long)ftrace_caller);
828#ifdef CONFIG_DYNAMIC_FTRACE_WITH_REGS
829 mod->arch.tramp_regs = create_ftrace_stub(sechdrs, mod,
830 (unsigned long)ftrace_regs_caller);
831 if (!mod->arch.tramp_regs)
832 return -ENOENT;
833#endif
834
835 if (!mod->arch.tramp)
836 return -ENOENT;
837
838 return 0;
839}
840#endif