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 <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