This commit was manufactured by cvs2svn to create branch 'vserver'.
[linux-2.6.git] / arch / ppc64 / kernel / vdso.c
1 /*
2  *  linux/arch/ppc64/kernel/vdso.c
3  *
4  *    Copyright (C) 2004 Benjamin Herrenschmidt, IBM Corp.
5  *                       <benh@kernel.crashing.org>
6  *
7  *  This program is free software; you can redistribute it and/or
8  *  modify it under the terms of the GNU General Public License
9  *  as published by the Free Software Foundation; either version
10  *  2 of the License, or (at your option) any later version.
11  */
12
13 #include <linux/config.h>
14 #include <linux/module.h>
15 #include <linux/errno.h>
16 #include <linux/sched.h>
17 #include <linux/kernel.h>
18 #include <linux/mm.h>
19 #include <linux/smp.h>
20 #include <linux/smp_lock.h>
21 #include <linux/stddef.h>
22 #include <linux/unistd.h>
23 #include <linux/slab.h>
24 #include <linux/user.h>
25 #include <linux/elf.h>
26 #include <linux/security.h>
27 #include <linux/bootmem.h>
28
29 #include <asm/pgtable.h>
30 #include <asm/system.h>
31 #include <asm/processor.h>
32 #include <asm/mmu.h>
33 #include <asm/mmu_context.h>
34 #include <asm/machdep.h>
35 #include <asm/cputable.h>
36 #include <asm/sections.h>
37 #include <asm/vdso.h>
38
39 #undef DEBUG
40
41 #ifdef DEBUG
42 #define DBG(fmt...) printk(fmt)
43 #else
44 #define DBG(fmt...)
45 #endif
46
47
48 /*
49  * The vDSOs themselves are here
50  */
51 extern char vdso64_start, vdso64_end;
52 extern char vdso32_start, vdso32_end;
53
54 static void *vdso64_kbase = &vdso64_start;
55 static void *vdso32_kbase = &vdso32_start;
56
57 unsigned int vdso64_pages;
58 unsigned int vdso32_pages;
59
60 /* Signal trampolines user addresses */
61
62 unsigned long vdso64_rt_sigtramp;
63 unsigned long vdso32_sigtramp;
64 unsigned long vdso32_rt_sigtramp;
65
66 /* Format of the patch table */
67 struct vdso_patch_def
68 {
69         u32             pvr_mask, pvr_value;
70         const char      *gen_name;
71         const char      *fix_name;
72 };
73
74 /* Table of functions to patch based on the CPU type/revision
75  *
76  * TODO: Improve by adding whole lists for each entry
77  */
78 static struct vdso_patch_def vdso_patches[] = {
79         {
80                 0xffff0000, 0x003a0000,         /* POWER5 */
81                 "__kernel_sync_dicache", "__kernel_sync_dicache_p5"
82         },
83         {
84                 0xffff0000, 0x003b0000,         /* POWER5 */
85                 "__kernel_sync_dicache", "__kernel_sync_dicache_p5"
86         },
87 };
88
89 /*
90  * Some infos carried around for each of them during parsing at
91  * boot time.
92  */
93 struct lib32_elfinfo
94 {
95         Elf32_Ehdr      *hdr;           /* ptr to ELF */
96         Elf32_Sym       *dynsym;        /* ptr to .dynsym section */
97         unsigned long   dynsymsize;     /* size of .dynsym section */
98         char            *dynstr;        /* ptr to .dynstr section */
99         unsigned long   text;           /* offset of .text section in .so */
100 };
101
102 struct lib64_elfinfo
103 {
104         Elf64_Ehdr      *hdr;
105         Elf64_Sym       *dynsym;
106         unsigned long   dynsymsize;
107         char            *dynstr;
108         unsigned long   text;
109 };
110
111
112 #ifdef __DEBUG
113 static void dump_one_vdso_page(struct page *pg, struct page *upg)
114 {
115         printk("kpg: %p (c:%d,f:%08lx)", __va(page_to_pfn(pg) << PAGE_SHIFT),
116                page_count(pg),
117                pg->flags);
118         if (upg/* && pg != upg*/) {
119                 printk(" upg: %p (c:%d,f:%08lx)", __va(page_to_pfn(upg) << PAGE_SHIFT),
120                        page_count(upg),
121                        upg->flags);
122         }
123         printk("\n");
124 }
125
126 static void dump_vdso_pages(struct vm_area_struct * vma)
127 {
128         int i;
129
130         if (!vma || test_thread_flag(TIF_32BIT)) {
131                 printk("vDSO32 @ %016lx:\n", (unsigned long)vdso32_kbase);
132                 for (i=0; i<vdso32_pages; i++) {
133                         struct page *pg = virt_to_page(vdso32_kbase + i*PAGE_SIZE);
134                         struct page *upg = (vma && vma->vm_mm) ?
135                                 follow_page(vma->vm_mm, vma->vm_start + i*PAGE_SIZE, 0)
136                                 : NULL;
137                         dump_one_vdso_page(pg, upg);
138                 }
139         }
140         if (!vma || !test_thread_flag(TIF_32BIT)) {
141                 printk("vDSO64 @ %016lx:\n", (unsigned long)vdso64_kbase);
142                 for (i=0; i<vdso64_pages; i++) {
143                         struct page *pg = virt_to_page(vdso64_kbase + i*PAGE_SIZE);
144                         struct page *upg = (vma && vma->vm_mm) ?
145                                 follow_page(vma->vm_mm, vma->vm_start + i*PAGE_SIZE, 0)
146                                 : NULL;
147                         dump_one_vdso_page(pg, upg);
148                 }
149         }
150 }
151 #endif /* DEBUG */
152
153 /*
154  * Keep a dummy vma_close for now, it will prevent VMA merging.
155  */
156 static void vdso_vma_close(struct vm_area_struct * vma)
157 {
158 }
159
160 /*
161  * Our nopage() function, maps in the actual vDSO kernel pages, they will
162  * be mapped read-only by do_no_page(), and eventually COW'ed, either
163  * right away for an initial write access, or by do_wp_page().
164  */
165 static struct page * vdso_vma_nopage(struct vm_area_struct * vma,
166                                      unsigned long address, int *type)
167 {
168         unsigned long offset = address - vma->vm_start;
169         struct page *pg;
170         void *vbase = test_thread_flag(TIF_32BIT) ? vdso32_kbase : vdso64_kbase;
171
172         DBG("vdso_vma_nopage(current: %s, address: %016lx, off: %lx)\n",
173             current->comm, address, offset);
174
175         if (address < vma->vm_start || address > vma->vm_end)
176                 return NOPAGE_SIGBUS;
177
178         /*
179          * Last page is systemcfg, special handling here, no get_page() a
180          * this is a reserved page
181          */
182         if ((vma->vm_end - address) <= PAGE_SIZE)
183                 return virt_to_page(systemcfg);
184
185         pg = virt_to_page(vbase + offset);
186         get_page(pg);
187         DBG(" ->page count: %d\n", page_count(pg));
188
189         return pg;
190 }
191
192 static struct vm_operations_struct vdso_vmops = {
193         .close  = vdso_vma_close,
194         .nopage = vdso_vma_nopage,
195 };
196
197 /*
198  * This is called from binfmt_elf, we create the special vma for the
199  * vDSO and insert it into the mm struct tree
200  */
201 int arch_setup_additional_pages(struct linux_binprm *bprm, int executable_stack)
202 {
203         struct mm_struct *mm = current->mm;
204         struct vm_area_struct *vma;
205         unsigned long vdso_pages;
206         unsigned long vdso_base;
207
208         if (test_thread_flag(TIF_32BIT)) {
209                 vdso_pages = vdso32_pages;
210                 vdso_base = VDSO32_MBASE;
211         } else {
212                 vdso_pages = vdso64_pages;
213                 vdso_base = VDSO64_MBASE;
214         }
215
216         current->thread.vdso_base = 0;
217
218         /* vDSO has a problem and was disabled, just don't "enable" it for the
219          * process
220          */
221         if (vdso_pages == 0)
222                 return 0;
223
224         vma = kmem_cache_alloc(vm_area_cachep, SLAB_KERNEL);
225         if (vma == NULL)
226                 return -ENOMEM;
227         if (security_vm_enough_memory(vdso_pages)) {
228                 kmem_cache_free(vm_area_cachep, vma);
229                 return -ENOMEM;
230         }
231         memset(vma, 0, sizeof(*vma));
232
233         /*
234          * pick a base address for the vDSO in process space. We try to put it
235          * at vdso_base which is the "natural" base for it, but we might fail
236          * and end up putting it elsewhere.
237          */
238         vdso_base = get_unmapped_area(NULL, vdso_base,
239                                       vdso_pages << PAGE_SHIFT, 0, 0);
240         if (vdso_base & ~PAGE_MASK)
241                 return (int)vdso_base;
242
243         current->thread.vdso_base = vdso_base;
244
245         vma->vm_mm = mm;
246         vma->vm_start = current->thread.vdso_base;
247
248         /*
249          * the VMA size is one page more than the vDSO since systemcfg
250          * is mapped in the last one
251          */
252         vma->vm_end = vma->vm_start + ((vdso_pages + 1) << PAGE_SHIFT);
253
254         /*
255          * our vma flags don't have VM_WRITE so by default, the process isn't allowed
256          * to write those pages.
257          * gdb can break that with ptrace interface, and thus trigger COW on those
258          * pages but it's then your responsibility to never do that on the "data" page
259          * of the vDSO or you'll stop getting kernel updates and your nice userland
260          * gettimeofday will be totally dead. It's fine to use that for setting
261          * breakpoints in the vDSO code pages though
262          */
263         vma->vm_flags = VM_READ | VM_EXEC | VM_MAYREAD | VM_MAYWRITE | VM_MAYEXEC;
264         vma->vm_flags |= mm->def_flags;
265         vma->vm_page_prot = protection_map[vma->vm_flags & 0x7];
266         vma->vm_ops = &vdso_vmops;
267
268         down_write(&mm->mmap_sem);
269         insert_vm_struct(mm, vma);
270         mm->total_vm += (vma->vm_end - vma->vm_start) >> PAGE_SHIFT;
271         up_write(&mm->mmap_sem);
272
273         return 0;
274 }
275
276 static void * __init find_section32(Elf32_Ehdr *ehdr, const char *secname,
277                                   unsigned long *size)
278 {
279         Elf32_Shdr *sechdrs;
280         unsigned int i;
281         char *secnames;
282
283         /* Grab section headers and strings so we can tell who is who */
284         sechdrs = (void *)ehdr + ehdr->e_shoff;
285         secnames = (void *)ehdr + sechdrs[ehdr->e_shstrndx].sh_offset;
286
287         /* Find the section they want */
288         for (i = 1; i < ehdr->e_shnum; i++) {
289                 if (strcmp(secnames+sechdrs[i].sh_name, secname) == 0) {
290                         if (size)
291                                 *size = sechdrs[i].sh_size;
292                         return (void *)ehdr + sechdrs[i].sh_offset;
293                 }
294         }
295         *size = 0;
296         return NULL;
297 }
298
299 static void * __init find_section64(Elf64_Ehdr *ehdr, const char *secname,
300                                   unsigned long *size)
301 {
302         Elf64_Shdr *sechdrs;
303         unsigned int i;
304         char *secnames;
305
306         /* Grab section headers and strings so we can tell who is who */
307         sechdrs = (void *)ehdr + ehdr->e_shoff;
308         secnames = (void *)ehdr + sechdrs[ehdr->e_shstrndx].sh_offset;
309
310         /* Find the section they want */
311         for (i = 1; i < ehdr->e_shnum; i++) {
312                 if (strcmp(secnames+sechdrs[i].sh_name, secname) == 0) {
313                         if (size)
314                                 *size = sechdrs[i].sh_size;
315                         return (void *)ehdr + sechdrs[i].sh_offset;
316                 }
317         }
318         if (size)
319                 *size = 0;
320         return NULL;
321 }
322
323 static Elf32_Sym * __init find_symbol32(struct lib32_elfinfo *lib, const char *symname)
324 {
325         unsigned int i;
326         char name[32], *c;
327
328         for (i = 0; i < (lib->dynsymsize / sizeof(Elf32_Sym)); i++) {
329                 if (lib->dynsym[i].st_name == 0)
330                         continue;
331                 strlcpy(name, lib->dynstr + lib->dynsym[i].st_name, 32);
332                 c = strchr(name, '@');
333                 if (c)
334                         *c = 0;
335                 if (strcmp(symname, name) == 0)
336                         return &lib->dynsym[i];
337         }
338         return NULL;
339 }
340
341 static Elf64_Sym * __init find_symbol64(struct lib64_elfinfo *lib, const char *symname)
342 {
343         unsigned int i;
344         char name[32], *c;
345
346         for (i = 0; i < (lib->dynsymsize / sizeof(Elf64_Sym)); i++) {
347                 if (lib->dynsym[i].st_name == 0)
348                         continue;
349                 strlcpy(name, lib->dynstr + lib->dynsym[i].st_name, 32);
350                 c = strchr(name, '@');
351                 if (c)
352                         *c = 0;
353                 if (strcmp(symname, name) == 0)
354                         return &lib->dynsym[i];
355         }
356         return NULL;
357 }
358
359 /* Note that we assume the section is .text and the symbol is relative to
360  * the library base
361  */
362 static unsigned long __init find_function32(struct lib32_elfinfo *lib, const char *symname)
363 {
364         Elf32_Sym *sym = find_symbol32(lib, symname);
365
366         if (sym == NULL) {
367                 printk(KERN_WARNING "vDSO32: function %s not found !\n", symname);
368                 return 0;
369         }
370         return sym->st_value - VDSO32_LBASE;
371 }
372
373 /* Note that we assume the section is .text and the symbol is relative to
374  * the library base
375  */
376 static unsigned long __init find_function64(struct lib64_elfinfo *lib, const char *symname)
377 {
378         Elf64_Sym *sym = find_symbol64(lib, symname);
379
380         if (sym == NULL) {
381                 printk(KERN_WARNING "vDSO64: function %s not found !\n", symname);
382                 return 0;
383         }
384 #ifdef VDS64_HAS_DESCRIPTORS
385         return *((u64 *)(vdso64_kbase + sym->st_value - VDSO64_LBASE)) - VDSO64_LBASE;
386 #else
387         return sym->st_value - VDSO64_LBASE;
388 #endif
389 }
390
391
392 static __init int vdso_do_find_sections(struct lib32_elfinfo *v32,
393                                         struct lib64_elfinfo *v64)
394 {
395         void *sect;
396
397         /*
398          * Locate symbol tables & text section
399          */
400
401         v32->dynsym = find_section32(v32->hdr, ".dynsym", &v32->dynsymsize);
402         v32->dynstr = find_section32(v32->hdr, ".dynstr", NULL);
403         if (v32->dynsym == NULL || v32->dynstr == NULL) {
404                 printk(KERN_ERR "vDSO32: a required symbol section was not found\n");
405                 return -1;
406         }
407         sect = find_section32(v32->hdr, ".text", NULL);
408         if (sect == NULL) {
409                 printk(KERN_ERR "vDSO32: the .text section was not found\n");
410                 return -1;
411         }
412         v32->text = sect - vdso32_kbase;
413
414         v64->dynsym = find_section64(v64->hdr, ".dynsym", &v64->dynsymsize);
415         v64->dynstr = find_section64(v64->hdr, ".dynstr", NULL);
416         if (v64->dynsym == NULL || v64->dynstr == NULL) {
417                 printk(KERN_ERR "vDSO64: a required symbol section was not found\n");
418                 return -1;
419         }
420         sect = find_section64(v64->hdr, ".text", NULL);
421         if (sect == NULL) {
422                 printk(KERN_ERR "vDSO64: the .text section was not found\n");
423                 return -1;
424         }
425         v64->text = sect - vdso64_kbase;
426
427         return 0;
428 }
429
430 static __init void vdso_setup_trampolines(struct lib32_elfinfo *v32,
431                                           struct lib64_elfinfo *v64)
432 {
433         /*
434          * Find signal trampolines
435          */
436
437         vdso64_rt_sigtramp      = find_function64(v64, "__kernel_sigtramp_rt64");
438         vdso32_sigtramp         = find_function32(v32, "__kernel_sigtramp32");
439         vdso32_rt_sigtramp      = find_function32(v32, "__kernel_sigtramp_rt32");
440 }
441
442 static __init int vdso_fixup_datapage(struct lib32_elfinfo *v32,
443                                        struct lib64_elfinfo *v64)
444 {
445         Elf32_Sym *sym32;
446         Elf64_Sym *sym64;
447
448         sym32 = find_symbol32(v32, "__kernel_datapage_offset");
449         if (sym32 == NULL) {
450                 printk(KERN_ERR "vDSO32: Can't find symbol __kernel_datapage_offset !\n");
451                 return -1;
452         }
453         *((int *)(vdso32_kbase + (sym32->st_value - VDSO32_LBASE))) =
454                 (vdso32_pages << PAGE_SHIFT) - (sym32->st_value - VDSO32_LBASE);
455
456         sym64 = find_symbol64(v64, "__kernel_datapage_offset");
457         if (sym64 == NULL) {
458                 printk(KERN_ERR "vDSO64: Can't find symbol __kernel_datapage_offset !\n");
459                 return -1;
460         }
461         *((int *)(vdso64_kbase + sym64->st_value - VDSO64_LBASE)) =
462                 (vdso64_pages << PAGE_SHIFT) - (sym64->st_value - VDSO64_LBASE);
463
464         return 0;
465 }
466
467 static int vdso_do_func_patch32(struct lib32_elfinfo *v32,
468                                 struct lib64_elfinfo *v64,
469                                 const char *orig, const char *fix)
470 {
471         Elf32_Sym *sym32_gen, *sym32_fix;
472
473         sym32_gen = find_symbol32(v32, orig);
474         if (sym32_gen == NULL) {
475                 printk(KERN_ERR "vDSO32: Can't find symbol %s !\n", orig);
476                 return -1;
477         }
478         sym32_fix = find_symbol32(v32, fix);
479         if (sym32_fix == NULL) {
480                 printk(KERN_ERR "vDSO32: Can't find symbol %s !\n", fix);
481                 return -1;
482         }
483         sym32_gen->st_value = sym32_fix->st_value;
484         sym32_gen->st_size = sym32_fix->st_size;
485         sym32_gen->st_info = sym32_fix->st_info;
486         sym32_gen->st_other = sym32_fix->st_other;
487         sym32_gen->st_shndx = sym32_fix->st_shndx;
488
489         return 0;
490 }
491
492 static int vdso_do_func_patch64(struct lib32_elfinfo *v32,
493                                 struct lib64_elfinfo *v64,
494                                 const char *orig, const char *fix)
495 {
496         Elf64_Sym *sym64_gen, *sym64_fix;
497
498         sym64_gen = find_symbol64(v64, orig);
499         if (sym64_gen == NULL) {
500                 printk(KERN_ERR "vDSO64: Can't find symbol %s !\n", orig);
501                 return -1;
502         }
503         sym64_fix = find_symbol64(v64, fix);
504         if (sym64_fix == NULL) {
505                 printk(KERN_ERR "vDSO64: Can't find symbol %s !\n", fix);
506                 return -1;
507         }
508         sym64_gen->st_value = sym64_fix->st_value;
509         sym64_gen->st_size = sym64_fix->st_size;
510         sym64_gen->st_info = sym64_fix->st_info;
511         sym64_gen->st_other = sym64_fix->st_other;
512         sym64_gen->st_shndx = sym64_fix->st_shndx;
513
514         return 0;
515 }
516
517 static __init int vdso_fixup_alt_funcs(struct lib32_elfinfo *v32,
518                                        struct lib64_elfinfo *v64)
519 {
520         u32 pvr;
521         int i;
522
523         pvr = mfspr(SPRN_PVR);
524         for (i = 0; i < ARRAY_SIZE(vdso_patches); i++) {
525                 struct vdso_patch_def *patch = &vdso_patches[i];
526                 int match = (pvr & patch->pvr_mask) == patch->pvr_value;
527
528                 DBG("patch %d (mask: %x, pvr: %x) : %s\n",
529                     i, patch->pvr_mask, patch->pvr_value, match ? "match" : "skip");
530
531                 if (!match)
532                         continue;
533
534                 DBG("replacing %s with %s...\n", patch->gen_name, patch->fix_name);
535
536                 /*
537                  * Patch the 32 bits and 64 bits symbols. Note that we do not patch
538                  * the "." symbol on 64 bits. It would be easy to do, but doesn't
539                  * seem to be necessary, patching the OPD symbol is enough.
540                  */
541                 vdso_do_func_patch32(v32, v64, patch->gen_name, patch->fix_name);
542                 vdso_do_func_patch64(v32, v64, patch->gen_name, patch->fix_name);
543         }
544
545         return 0;
546 }
547
548
549 static __init int vdso_setup(void)
550 {
551         struct lib32_elfinfo    v32;
552         struct lib64_elfinfo    v64;
553
554         v32.hdr = vdso32_kbase;
555         v64.hdr = vdso64_kbase;
556
557         if (vdso_do_find_sections(&v32, &v64))
558                 return -1;
559
560         if (vdso_fixup_datapage(&v32, &v64))
561                 return -1;
562
563         if (vdso_fixup_alt_funcs(&v32, &v64))
564                 return -1;
565
566         vdso_setup_trampolines(&v32, &v64);
567
568         return 0;
569 }
570
571 void __init vdso_init(void)
572 {
573         int i;
574
575         vdso64_pages = (&vdso64_end - &vdso64_start) >> PAGE_SHIFT;
576         vdso32_pages = (&vdso32_end - &vdso32_start) >> PAGE_SHIFT;
577
578         DBG("vdso64_kbase: %p, 0x%x pages, vdso32_kbase: %p, 0x%x pages\n",
579                vdso64_kbase, vdso64_pages, vdso32_kbase, vdso32_pages);
580
581         /*
582          * Initialize the vDSO images in memory, that is do necessary
583          * fixups of vDSO symbols, locate trampolines, etc...
584          */
585         if (vdso_setup()) {
586                 printk(KERN_ERR "vDSO setup failure, not enabled !\n");
587                 /* XXX should free pages here ? */
588                 vdso64_pages = vdso32_pages = 0;
589                 return;
590         }
591
592         /* Make sure pages are in the correct state */
593         for (i = 0; i < vdso64_pages; i++) {
594                 struct page *pg = virt_to_page(vdso64_kbase + i*PAGE_SIZE);
595                 ClearPageReserved(pg);
596                 get_page(pg);
597         }
598         for (i = 0; i < vdso32_pages; i++) {
599                 struct page *pg = virt_to_page(vdso32_kbase + i*PAGE_SIZE);
600                 ClearPageReserved(pg);
601                 get_page(pg);
602         }
603 }
604
605 int in_gate_area_no_task(unsigned long addr)
606 {
607         return 0;
608 }
609
610 int in_gate_area(struct task_struct *task, unsigned long addr)
611 {
612         return 0;
613 }
614
615 struct vm_area_struct *get_gate_vma(struct task_struct *tsk)
616 {
617         return NULL;
618 }
619