target/linux/generic/pending-4.9/305-mips_module_reloc.patch

   1 From: Felix Fietkau <nbd@nbd.name>
   2 Subject: mips: replace -mlong-calls with -mno-long-calls to make function calls faster in kernel modules to achieve this, try to
   3
   4 lede-commit: 3b3d64743ba2a874df9d70cd19e242205b0a788c
   5 Signed-off-by: Felix Fietkau <nbd@nbd.name>
   6 ---
   7  arch/mips/Makefile             |   5 +
   8  arch/mips/include/asm/module.h |   5 +
   9  arch/mips/kernel/module.c      | 279 ++++++++++++++++++++++++++++++++++++++++-
  10  3 files changed, 284 insertions(+), 5 deletions(-)
  11
  12 --- a/arch/mips/Makefile
  13 +++ b/arch/mips/Makefile
  14 @@ -93,8 +93,13 @@ all-$(CONFIG_SYS_SUPPORTS_ZBOOT)+= vmlin
  15  cflags-y                       += -G 0 -mno-abicalls -fno-pic -pipe -mno-branch-likely
  16  cflags-y                       += -msoft-float
  17  LDFLAGS_vmlinux                        += -G 0 -static -n -nostdlib
  18 +ifdef CONFIG_64BIT
  19  KBUILD_AFLAGS_MODULE           += -mlong-calls
  20  KBUILD_CFLAGS_MODULE           += -mlong-calls
  21 +else
  22 +KBUILD_AFLAGS_MODULE           += -mno-long-calls
  23 +KBUILD_CFLAGS_MODULE           += -mno-long-calls
  24 +endif
  25
  26  ifeq ($(CONFIG_RELOCATABLE),y)
  27  LDFLAGS_vmlinux                        += --emit-relocs
  28 --- a/arch/mips/include/asm/module.h
  29 +++ b/arch/mips/include/asm/module.h
  30 @@ -11,6 +11,11 @@ struct mod_arch_specific {
  31         const struct exception_table_entry *dbe_start;
  32         const struct exception_table_entry *dbe_end;
  33         struct mips_hi16 *r_mips_hi16_list;
  34 +
  35 +       void *phys_plt_tbl;
  36 +       void *virt_plt_tbl;
  37 +       unsigned int phys_plt_offset;
  38 +       unsigned int virt_plt_offset;
  39  };
  40
  41  typedef uint8_t Elf64_Byte;            /* Type for a 8-bit quantity.  */
  42 --- a/arch/mips/kernel/module.c
  43 +++ b/arch/mips/kernel/module.c
  44 @@ -44,14 +44,221 @@ struct mips_hi16 {
  45  static LIST_HEAD(dbe_list);
  46  static DEFINE_SPINLOCK(dbe_lock);
  47
  48 -#ifdef MODULE_START
  49 +/*
  50 + * Get the potential max trampolines size required of the init and
  51 + * non-init sections. Only used if we cannot find enough contiguous
  52 + * physically mapped memory to put the module into.
  53 + */
  54 +static unsigned int
  55 +get_plt_size(const Elf_Ehdr *hdr, const Elf_Shdr *sechdrs,
  56 +             const char *secstrings, unsigned int symindex, bool is_init)
  57 +{
  58 +       unsigned long ret = 0;
  59 +       unsigned int i, j;
  60 +       Elf_Sym *syms;
  61 +
  62 +       /* Everything marked ALLOC (this includes the exported symbols) */
  63 +       for (i = 1; i < hdr->e_shnum; ++i) {
  64 +               unsigned int info = sechdrs[i].sh_info;
  65 +
  66 +               if (sechdrs[i].sh_type != SHT_REL
  67 +                   && sechdrs[i].sh_type != SHT_RELA)
  68 +                       continue;
  69 +
  70 +               /* Not a valid relocation section? */
  71 +               if (info >= hdr->e_shnum)
  72 +                       continue;
  73 +
  74 +               /* Don't bother with non-allocated sections */
  75 +               if (!(sechdrs[info].sh_flags & SHF_ALLOC))
  76 +                       continue;
  77 +
  78 +               /* If it's called *.init*, and we're not init, we're
  79 +                   not interested */
  80 +               if ((strstr(secstrings + sechdrs[i].sh_name, ".init") != 0)
  81 +                   != is_init)
  82 +                       continue;
  83 +
  84 +               syms = (Elf_Sym *) sechdrs[symindex].sh_addr;
  85 +               if (sechdrs[i].sh_type == SHT_REL) {
  86 +                       Elf_Mips_Rel *rel = (void *) sechdrs[i].sh_addr;
  87 +                       unsigned int size = sechdrs[i].sh_size / sizeof(*rel);
  88 +
  89 +                       for (j = 0; j < size; ++j) {
  90 +                               Elf_Sym *sym;
  91 +
  92 +                               if (ELF_MIPS_R_TYPE(rel[j]) != R_MIPS_26)
  93 +                                       continue;
  94 +
  95 +                               sym = syms + ELF_MIPS_R_SYM(rel[j]);
  96 +                               if (!is_init && sym->st_shndx != SHN_UNDEF)
  97 +                                       continue;
  98 +
  99 +                               ret += 4 * sizeof(int);
 100 +                       }
 101 +               } else {
 102 +                       Elf_Mips_Rela *rela = (void *) sechdrs[i].sh_addr;
 103 +                       unsigned int size = sechdrs[i].sh_size / sizeof(*rela);
 104 +
 105 +                       for (j = 0; j < size; ++j) {
 106 +                               Elf_Sym *sym;
 107 +
 108 +                               if (ELF_MIPS_R_TYPE(rela[j]) != R_MIPS_26)
 109 +                                       continue;
 110 +
 111 +                               sym = syms + ELF_MIPS_R_SYM(rela[j]);
 112 +                               if (!is_init && sym->st_shndx != SHN_UNDEF)
 113 +                                       continue;
 114 +
 115 +                               ret += 4 * sizeof(int);
 116 +                       }
 117 +               }
 118 +       }
 119 +
 120 +       return ret;
 121 +}
 122 +
 123 +#ifndef MODULE_START
 124 +static void *alloc_phys(unsigned long size)
 125 +{
 126 +       unsigned order;
 127 +       struct page *page;
 128 +       struct page *p;
 129 +
 130 +       size = PAGE_ALIGN(size);
 131 +       order = get_order(size);
 132 +
 133 +       page = alloc_pages(GFP_KERNEL | __GFP_NORETRY | __GFP_NOWARN |
 134 +                       __GFP_THISNODE, order);
 135 +       if (!page)
 136 +               return NULL;
 137 +
 138 +       split_page(page, order);
 139 +
 140 +       /* mark all pages except for the last one */
 141 +       for (p = page; p + 1 < page + (size >> PAGE_SHIFT); ++p)
 142 +               set_bit(PG_owner_priv_1, &p->flags);
 143 +
 144 +       for (p = page + (size >> PAGE_SHIFT); p < page + (1 << order); ++p)
 145 +               __free_page(p);
 146 +
 147 +       return page_address(page);
 148 +}
 149 +#endif
 150 +
 151 +static void free_phys(void *ptr)
 152 +{
 153 +       struct page *page;
 154 +       bool free;
 155 +
 156 +       page = virt_to_page(ptr);
 157 +       do {
 158 +               free = test_and_clear_bit(PG_owner_priv_1, &page->flags);
 159 +               __free_page(page);
 160 +               page++;
 161 +       } while (free);
 162 +}
 163 +
 164 +
 165  void *module_alloc(unsigned long size)
 166  {
 167 +#ifdef MODULE_START
 168         return __vmalloc_node_range(size, 1, MODULE_START, MODULE_END,
 169                                 GFP_KERNEL, PAGE_KERNEL, 0, NUMA_NO_NODE,
 170                                 __builtin_return_address(0));
 171 +#else
 172 +       void *ptr;
 173 +
 174 +       if (size == 0)
 175 +               return NULL;
 176 +
 177 +       ptr = alloc_phys(size);
 178 +
 179 +       /* If we failed to allocate physically contiguous memory,
 180 +        * fall back to regular vmalloc. The module loader code will
 181 +        * create jump tables to handle long jumps */
 182 +       if (!ptr)
 183 +               return vmalloc(size);
 184 +
 185 +       return ptr;
 186 +#endif
 187  }
 188 +
 189 +static inline bool is_phys_addr(void *ptr)
 190 +{
 191 +#ifdef CONFIG_64BIT
 192 +       return (KSEGX((unsigned long)ptr) == CKSEG0);
 193 +#else
 194 +       return (KSEGX(ptr) == KSEG0);
 195  #endif
 196 +}
 197 +
 198 +/* Free memory returned from module_alloc */
 199 +void module_memfree(void *module_region)
 200 +{
 201 +       if (is_phys_addr(module_region))
 202 +               free_phys(module_region);
 203 +       else
 204 +               vfree(module_region);
 205 +}
 206 +
 207 +static void *__module_alloc(int size, bool phys)
 208 +{
 209 +       void *ptr;
 210 +
 211 +       if (phys)
 212 +               ptr = kmalloc(size, GFP_KERNEL);
 213 +       else
 214 +               ptr = vmalloc(size);
 215 +       return ptr;
 216 +}
 217 +
 218 +static void __module_free(void *ptr)
 219 +{
 220 +       if (is_phys_addr(ptr))
 221 +               kfree(ptr);
 222 +       else
 223 +               vfree(ptr);
 224 +}
 225 +
 226 +int module_frob_arch_sections(Elf_Ehdr *hdr, Elf_Shdr *sechdrs,
 227 +                             char *secstrings, struct module *mod)
 228 +{
 229 +       unsigned int symindex = 0;
 230 +       unsigned int core_size, init_size;
 231 +       int i;
 232 +
 233 +       mod->arch.phys_plt_offset = 0;
 234 +       mod->arch.virt_plt_offset = 0;
 235 +       mod->arch.phys_plt_tbl = NULL;
 236 +       mod->arch.virt_plt_tbl = NULL;
 237 +
 238 +       if (IS_ENABLED(CONFIG_64BIT))
 239 +               return 0;
 240 +
 241 +       for (i = 1; i < hdr->e_shnum; i++)
 242 +               if (sechdrs[i].sh_type == SHT_SYMTAB)
 243 +                       symindex = i;
 244 +
 245 +       core_size = get_plt_size(hdr, sechdrs, secstrings, symindex, false);
 246 +       init_size = get_plt_size(hdr, sechdrs, secstrings, symindex, true);
 247 +
 248 +       if ((core_size + init_size) == 0)
 249 +               return 0;
 250 +
 251 +       mod->arch.phys_plt_tbl = __module_alloc(core_size + init_size, 1);
 252 +       if (!mod->arch.phys_plt_tbl)
 253 +               return -ENOMEM;
 254 +
 255 +       mod->arch.virt_plt_tbl = __module_alloc(core_size + init_size, 0);
 256 +       if (!mod->arch.virt_plt_tbl) {
 257 +               __module_free(mod->arch.phys_plt_tbl);
 258 +               mod->arch.phys_plt_tbl = NULL;
 259 +               return -ENOMEM;
 260 +       }
 261 +
 262 +       return 0;
 263 +}
 264
 265  int apply_r_mips_none(struct module *me, u32 *location, Elf_Addr v)
 266  {
 267 @@ -65,8 +272,39 @@ static int apply_r_mips_32_rel(struct mo
 268         return 0;
 269  }
 270
 271 +static Elf_Addr add_plt_entry_to(unsigned *plt_offset,
 272 +                                void *start, Elf_Addr v)
 273 +{
 274 +       unsigned *tramp = start + *plt_offset;
 275 +       *plt_offset += 4 * sizeof(int);
 276 +
 277 +       /* adjust carry for addiu */
 278 +       if (v & 0x00008000)
 279 +               v += 0x10000;
 280 +
 281 +       tramp[0] = 0x3c190000 | (v >> 16);      /* lui t9, hi16 */
 282 +       tramp[1] = 0x27390000 | (v & 0xffff);   /* addiu t9, t9, lo16 */
 283 +       tramp[2] = 0x03200008;                  /* jr t9 */
 284 +       tramp[3] = 0x00000000;                  /* nop */
 285 +
 286 +       return (Elf_Addr) tramp;
 287 +}
 288 +
 289 +static Elf_Addr add_plt_entry(struct module *me, void *location, Elf_Addr v)
 290 +{
 291 +       if (is_phys_addr(location))
 292 +               return add_plt_entry_to(&me->arch.phys_plt_offset,
 293 +                               me->arch.phys_plt_tbl, v);
 294 +       else
 295 +               return add_plt_entry_to(&me->arch.virt_plt_offset,
 296 +                               me->arch.virt_plt_tbl, v);
 297 +
 298 +}
 299 +
 300  static int apply_r_mips_26_rel(struct module *me, u32 *location, Elf_Addr v)
 301  {
 302 +       u32 ofs = *location & 0x03ffffff;
 303 +
 304         if (v % 4) {
 305                 pr_err("module %s: dangerous R_MIPS_26 REL relocation\n",
 306                        me->name);
 307 @@ -74,13 +312,17 @@ static int apply_r_mips_26_rel(struct mo
 308         }
 309
 310         if ((v & 0xf0000000) != (((unsigned long)location + 4) & 0xf0000000)) {
 311 -               pr_err("module %s: relocation overflow\n",
 312 -                      me->name);
 313 -               return -ENOEXEC;
 314 +               v = add_plt_entry(me, location, v + (ofs << 2));
 315 +               if (!v) {
 316 +                       pr_err("module %s: relocation overflow\n",
 317 +                              me->name);
 318 +                       return -ENOEXEC;
 319 +               }
 320 +               ofs = 0;
 321         }
 322
 323         *location = (*location & ~0x03ffffff) |
 324 -                   ((*location + (v >> 2)) & 0x03ffffff);
 325 +                   ((ofs + (v >> 2)) & 0x03ffffff);
 326
 327         return 0;
 328  }
 329 @@ -349,9 +591,36 @@ int module_finalize(const Elf_Ehdr *hdr,
 330                 list_add(&me->arch.dbe_list, &dbe_list);
 331                 spin_unlock_irq(&dbe_lock);
 332         }
 333 +
 334 +       /* Get rid of the fixup trampoline if we're running the module
 335 +        * from physically mapped address space */
 336 +       if (me->arch.phys_plt_offset == 0) {
 337 +               __module_free(me->arch.phys_plt_tbl);
 338 +               me->arch.phys_plt_tbl = NULL;
 339 +       }
 340 +       if (me->arch.virt_plt_offset == 0) {
 341 +               __module_free(me->arch.virt_plt_tbl);
 342 +               me->arch.virt_plt_tbl = NULL;
 343 +       }
 344 +
 345         return 0;
 346  }
 347
 348 +void module_arch_freeing_init(struct module *mod)
 349 +{
 350 +       if (mod->state == MODULE_STATE_LIVE)
 351 +               return;
 352 +
 353 +       if (mod->arch.phys_plt_tbl) {
 354 +               __module_free(mod->arch.phys_plt_tbl);
 355 +               mod->arch.phys_plt_tbl = NULL;
 356 +       }
 357 +       if (mod->arch.virt_plt_tbl) {
 358 +               __module_free(mod->arch.virt_plt_tbl);
 359 +               mod->arch.virt_plt_tbl = NULL;
 360 +       }
 361 +}
 362 +
 363  void module_arch_cleanup(struct module *mod)
 364  {
 365         spin_lock_irq(&dbe_lock);