arch/powerpc/oprofile/cell/vma_map.c

   1 /*
   2  * Cell Broadband Engine OProfile Support
   3  *
   4  * (C) Copyright IBM Corporation 2006
   5  *
   6  * Author: Maynard Johnson <maynardj@us.ibm.com>
   7  *
   8  * This program is free software; you can redistribute it and/or
   9  * modify it under the terms of the GNU General Public License
  10  * as published by the Free Software Foundation; either version
  11  * 2 of the License, or (at your option) any later version.
  12  */
  13
  14 /* The code in this source file is responsible for generating
  15  * vma-to-fileOffset maps for both overlay and non-overlay SPU
  16  * applications.
  17  */
  18
  19 #include <linux/mm.h>
  20 #include <linux/string.h>
  21 #include <linux/uaccess.h>
  22 #include <linux/elf.h>
  23 #include <linux/slab.h>
  24 #include "pr_util.h"
  25
  26
  27 void vma_map_free(struct vma_to_fileoffset_map *map)
  28 {
  29         while (map) {
  30                 struct vma_to_fileoffset_map *next = map->next;
  31                 kfree(map);
  32                 map = next;
  33         }
  34 }
  35
  36 unsigned int
  37 vma_map_lookup(struct vma_to_fileoffset_map *map, unsigned int vma,
  38                const struct spu *aSpu, int *grd_val)
  39 {
  40         /*
  41          * Default the offset to the physical address + a flag value.
  42          * Addresses of dynamically generated code can't be found in the vma
  43          * map.  For those addresses the flagged value will be sent on to
  44          * the user space tools so they can be reported rather than just
  45          * thrown away.
  46          */
  47         u32 offset = 0x10000000 + vma;
  48         u32 ovly_grd;
  49
  50         for (; map; map = map->next) {
  51                 if (vma < map->vma || vma >= map->vma + map->size)
  52                         continue;
  53
  54                 if (map->guard_ptr) {
  55                         ovly_grd = *(u32 *)(aSpu->local_store + map->guard_ptr);
  56                         if (ovly_grd != map->guard_val)
  57                                 continue;
  58                         *grd_val = ovly_grd;
  59                 }
  60                 offset = vma - map->vma + map->offset;
  61                 break;
  62         }
  63
  64         return offset;
  65 }
  66
  67 static struct vma_to_fileoffset_map *
  68 vma_map_add(struct vma_to_fileoffset_map *map, unsigned int vma,
  69             unsigned int size, unsigned int offset, unsigned int guard_ptr,
  70             unsigned int guard_val)
  71 {
  72         struct vma_to_fileoffset_map *new =
  73                 kzalloc(sizeof(struct vma_to_fileoffset_map), GFP_KERNEL);
  74         if (!new) {
  75                 printk(KERN_ERR "SPU_PROF: %s, line %d: malloc failed\n",
  76                        __func__, __LINE__);
  77                 vma_map_free(map);
  78                 return NULL;
  79         }
  80
  81         new->next = map;
  82         new->vma = vma;
  83         new->size = size;
  84         new->offset = offset;
  85         new->guard_ptr = guard_ptr;
  86         new->guard_val = guard_val;
  87
  88         return new;
  89 }
  90
  91
  92 /* Parse SPE ELF header and generate a list of vma_maps.
  93  * A pointer to the first vma_map in the generated list
  94  * of vma_maps is returned.  */
  95 struct vma_to_fileoffset_map *create_vma_map(const struct spu *aSpu,
  96                                              unsigned long __spu_elf_start)
  97 {
  98         static const unsigned char expected[EI_PAD] = {
  99                 [EI_MAG0] = ELFMAG0,
 100                 [EI_MAG1] = ELFMAG1,
 101                 [EI_MAG2] = ELFMAG2,
 102                 [EI_MAG3] = ELFMAG3,
 103                 [EI_CLASS] = ELFCLASS32,
 104                 [EI_DATA] = ELFDATA2MSB,
 105                 [EI_VERSION] = EV_CURRENT,
 106                 [EI_OSABI] = ELFOSABI_NONE
 107         };
 108
 109         int grd_val;
 110         struct vma_to_fileoffset_map *map = NULL;
 111         void __user *spu_elf_start = (void __user *)__spu_elf_start;
 112         struct spu_overlay_info ovly;
 113         unsigned int overlay_tbl_offset = -1;
 114         Elf32_Phdr __user *phdr_start;
 115         Elf32_Shdr __user *shdr_start;
 116         Elf32_Ehdr ehdr;
 117         Elf32_Phdr phdr;
 118         Elf32_Shdr shdr, shdr_str;
 119         Elf32_Sym sym;
 120         int i, j;
 121         char name[32];
 122
 123         unsigned int ovly_table_sym = 0;
 124         unsigned int ovly_buf_table_sym = 0;
 125         unsigned int ovly_table_end_sym = 0;
 126         unsigned int ovly_buf_table_end_sym = 0;
 127         struct spu_overlay_info __user *ovly_table;
 128         unsigned int n_ovlys;
 129
 130         /* Get and validate ELF header.  */
 131
 132         if (copy_from_user(&ehdr, spu_elf_start, sizeof (ehdr)))
 133                 goto fail;
 134
 135         if (memcmp(ehdr.e_ident, expected, EI_PAD) != 0) {
 136                 printk(KERN_ERR "SPU_PROF: "
 137                        "%s, line %d: Unexpected e_ident parsing SPU ELF\n",
 138                        __func__, __LINE__);
 139                 goto fail;
 140         }
 141         if (ehdr.e_machine != EM_SPU) {
 142                 printk(KERN_ERR "SPU_PROF: "
 143                        "%s, line %d: Unexpected e_machine parsing SPU ELF\n",
 144                        __func__,  __LINE__);
 145                 goto fail;
 146         }
 147         if (ehdr.e_type != ET_EXEC) {
 148                 printk(KERN_ERR "SPU_PROF: "
 149                        "%s, line %d: Unexpected e_type parsing SPU ELF\n",
 150                        __func__, __LINE__);
 151                 goto fail;
 152         }
 153         phdr_start = spu_elf_start + ehdr.e_phoff;
 154         shdr_start = spu_elf_start + ehdr.e_shoff;
 155
 156         /* Traverse program headers.  */
 157         for (i = 0; i < ehdr.e_phnum; i++) {
 158                 if (copy_from_user(&phdr, phdr_start + i, sizeof(phdr)))
 159                         goto fail;
 160
 161                 if (phdr.p_type != PT_LOAD)
 162                         continue;
 163                 if (phdr.p_flags & (1 << 27))
 164                         continue;
 165
 166                 map = vma_map_add(map, phdr.p_vaddr, phdr.p_memsz,
 167                                   phdr.p_offset, 0, 0);
 168                 if (!map)
 169                         goto fail;
 170         }
 171
 172         pr_debug("SPU_PROF: Created non-overlay maps\n");
 173         /* Traverse section table and search for overlay-related symbols.  */
 174         for (i = 0; i < ehdr.e_shnum; i++) {
 175                 if (copy_from_user(&shdr, shdr_start + i, sizeof(shdr)))
 176                         goto fail;
 177
 178                 if (shdr.sh_type != SHT_SYMTAB)
 179                         continue;
 180                 if (shdr.sh_entsize != sizeof (sym))
 181                         continue;
 182
 183                 if (copy_from_user(&shdr_str,
 184                                    shdr_start + shdr.sh_link,
 185                                    sizeof(shdr)))
 186                         goto fail;
 187
 188                 if (shdr_str.sh_type != SHT_STRTAB)
 189                         goto fail;
 190
 191                 for (j = 0; j < shdr.sh_size / sizeof (sym); j++) {
 192                         if (copy_from_user(&sym, spu_elf_start +
 193                                                  shdr.sh_offset +
 194                                                  j * sizeof (sym),
 195                                            sizeof (sym)))
 196                                 goto fail;
 197
 198                         if (copy_from_user(name,
 199                                            spu_elf_start + shdr_str.sh_offset +
 200                                            sym.st_name,
 201                                            20))
 202                                 goto fail;
 203
 204                         if (memcmp(name, "_ovly_table", 12) == 0)
 205                                 ovly_table_sym = sym.st_value;
 206                         if (memcmp(name, "_ovly_buf_table", 16) == 0)
 207                                 ovly_buf_table_sym = sym.st_value;
 208                         if (memcmp(name, "_ovly_table_end", 16) == 0)
 209                                 ovly_table_end_sym = sym.st_value;
 210                         if (memcmp(name, "_ovly_buf_table_end", 20) == 0)
 211                                 ovly_buf_table_end_sym = sym.st_value;
 212                 }
 213         }
 214
 215         /* If we don't have overlays, we're done.  */
 216         if (ovly_table_sym == 0 || ovly_buf_table_sym == 0
 217             || ovly_table_end_sym == 0 || ovly_buf_table_end_sym == 0) {
 218                 pr_debug("SPU_PROF: No overlay table found\n");
 219                 goto out;
 220         } else {
 221                 pr_debug("SPU_PROF: Overlay table found\n");
 222         }
 223
 224         /* The _ovly_table symbol represents a table with one entry
 225          * per overlay section.  The _ovly_buf_table symbol represents
 226          * a table with one entry per overlay region.
 227          * The struct spu_overlay_info gives the structure of the _ovly_table
 228          * entries.  The structure of _ovly_table_buf is simply one
 229          * u32 word per entry.
 230          */
 231         overlay_tbl_offset = vma_map_lookup(map, ovly_table_sym,
 232                                             aSpu, &grd_val);
 233         if (overlay_tbl_offset > 0x10000000) {
 234                 printk(KERN_ERR "SPU_PROF: "
 235                        "%s, line %d: Error finding SPU overlay table\n",
 236                        __func__, __LINE__);
 237                 goto fail;
 238         }
 239         ovly_table = spu_elf_start + overlay_tbl_offset;
 240
 241         n_ovlys = (ovly_table_end_sym -
 242                    ovly_table_sym) / sizeof (ovly);
 243
 244         /* Traverse overlay table.  */
 245         for (i = 0; i < n_ovlys; i++) {
 246                 if (copy_from_user(&ovly, ovly_table + i, sizeof (ovly)))
 247                         goto fail;
 248
 249                 /* The ovly.vma/size/offset arguments are analogous to the same
 250                  * arguments used above for non-overlay maps.  The final two
 251                  * args are referred to as the guard pointer and the guard
 252                  * value.
 253                  * The guard pointer is an entry in the _ovly_buf_table,
 254                  * computed using ovly.buf as the index into the table.  Since
 255                  * ovly.buf values begin at '1' to reference the first (or 0th)
 256                  * entry in the _ovly_buf_table, the computation subtracts 1
 257                  * from ovly.buf.
 258                  * The guard value is stored in the _ovly_buf_table entry and
 259                  * is an index (starting at 1) back to the _ovly_table entry
 260                  * that is pointing at this _ovly_buf_table entry.  So, for
 261                  * example, for an overlay scenario with one overlay segment
 262                  * and two overlay sections:
 263                  *      - Section 1 points to the first entry of the
 264                  *        _ovly_buf_table, which contains a guard value
 265                  *        of '1', referencing the first (index=0) entry of
 266                  *        _ovly_table.
 267                  *      - Section 2 points to the second entry of the
 268                  *        _ovly_buf_table, which contains a guard value
 269                  *        of '2', referencing the second (index=1) entry of
 270                  *        _ovly_table.
 271                  */
 272                 map = vma_map_add(map, ovly.vma, ovly.size, ovly.offset,
 273                                   ovly_buf_table_sym + (ovly.buf-1) * 4, i+1);
 274                 if (!map)
 275                         goto fail;
 276         }
 277         goto out;
 278
 279  fail:
 280         map = NULL;
 281  out:
 282         return map;
 283 }