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