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 diff --git a/arch/mips/Makefile b/arch/mips/Makefile
  13 index 48dc1a9c3e42..77bf5db20d65 100644
  14 --- a/arch/mips/Makefile
  15 +++ b/arch/mips/Makefile
  16 @@ -93,8 +93,13 @@ all-$(CONFIG_SYS_SUPPORTS_ZBOOT)+= vmlinuz
  17  cflags-y                       += -G 0 -mno-abicalls -fno-pic -pipe -mno-branch-likely
  18  cflags-y                       += -msoft-float
  19  LDFLAGS_vmlinux                        += -G 0 -static -n -nostdlib
  20 +ifdef CONFIG_64BIT
  21  KBUILD_AFLAGS_MODULE           += -mlong-calls
  22  KBUILD_CFLAGS_MODULE           += -mlong-calls
  23 +else
  24 +KBUILD_AFLAGS_MODULE           += -mno-long-calls
  25 +KBUILD_CFLAGS_MODULE           += -mno-long-calls
  26 +endif
  27
  28  ifeq ($(CONFIG_RELOCATABLE),y)
  29  LDFLAGS_vmlinux                        += --emit-relocs
  30 diff --git a/arch/mips/include/asm/module.h b/arch/mips/include/asm/module.h
  31 index 702c273e67a9..1d4f3b37cefe 100644
  32 --- a/arch/mips/include/asm/module.h
  33 +++ b/arch/mips/include/asm/module.h
  34 @@ -11,6 +11,11 @@ struct mod_arch_specific {
  35         const struct exception_table_entry *dbe_start;
  36         const struct exception_table_entry *dbe_end;
  37         struct mips_hi16 *r_mips_hi16_list;
  38 +
  39 +       void *phys_plt_tbl;
  40 +       void *virt_plt_tbl;
  41 +       unsigned int phys_plt_offset;
  42 +       unsigned int virt_plt_offset;
  43  };
  44
  45  typedef uint8_t Elf64_Byte;            /* Type for a 8-bit quantity.  */
  46 diff --git a/arch/mips/kernel/module.c b/arch/mips/kernel/module.c
  47 index 94627a3a6a0d..947981a9aa72 100644
  48 --- a/arch/mips/kernel/module.c
  49 +++ b/arch/mips/kernel/module.c
  50 @@ -44,14 +44,221 @@ struct mips_hi16 {
  51  static LIST_HEAD(dbe_list);
  52  static DEFINE_SPINLOCK(dbe_lock);
  53
  54 -#ifdef MODULE_START
  55 +/*
  56 + * Get the potential max trampolines size required of the init and
  57 + * non-init sections. Only used if we cannot find enough contiguous
  58 + * physically mapped memory to put the module into.
  59 + */
  60 +static unsigned int
  61 +get_plt_size(const Elf_Ehdr *hdr, const Elf_Shdr *sechdrs,
  62 +             const char *secstrings, unsigned int symindex, bool is_init)
  63 +{
  64 +       unsigned long ret = 0;
  65 +       unsigned int i, j;
  66 +       Elf_Sym *syms;
  67 +
  68 +       /* Everything marked ALLOC (this includes the exported symbols) */
  69 +       for (i = 1; i < hdr->e_shnum; ++i) {
  70 +               unsigned int info = sechdrs[i].sh_info;
  71 +
  72 +               if (sechdrs[i].sh_type != SHT_REL
  73 +                   && sechdrs[i].sh_type != SHT_RELA)
  74 +                       continue;
  75 +
  76 +               /* Not a valid relocation section? */
  77 +               if (info >= hdr->e_shnum)
  78 +                       continue;
  79 +
  80 +               /* Don't bother with non-allocated sections */
  81 +               if (!(sechdrs[info].sh_flags & SHF_ALLOC))
  82 +                       continue;
  83 +
  84 +               /* If it's called *.init*, and we're not init, we're
  85 +                   not interested */
  86 +               if ((strstr(secstrings + sechdrs[i].sh_name, ".init") != 0)
  87 +                   != is_init)
  88 +                       continue;
  89 +
  90 +               syms = (Elf_Sym *) sechdrs[symindex].sh_addr;
  91 +               if (sechdrs[i].sh_type == SHT_REL) {
  92 +                       Elf_Mips_Rel *rel = (void *) sechdrs[i].sh_addr;
  93 +                       unsigned int size = sechdrs[i].sh_size / sizeof(*rel);
  94 +
  95 +                       for (j = 0; j < size; ++j) {
  96 +                               Elf_Sym *sym;
  97 +
  98 +                               if (ELF_MIPS_R_TYPE(rel[j]) != R_MIPS_26)
  99 +                                       continue;
 100 +
 101 +                               sym = syms + ELF_MIPS_R_SYM(rel[j]);
 102 +                               if (!is_init && sym->st_shndx != SHN_UNDEF)
 103 +                                       continue;
 104 +
 105 +                               ret += 4 * sizeof(int);
 106 +                       }
 107 +               } else {
 108 +                       Elf_Mips_Rela *rela = (void *) sechdrs[i].sh_addr;
 109 +                       unsigned int size = sechdrs[i].sh_size / sizeof(*rela);
 110 +
 111 +                       for (j = 0; j < size; ++j) {
 112 +                               Elf_Sym *sym;
 113 +
 114 +                               if (ELF_MIPS_R_TYPE(rela[j]) != R_MIPS_26)
 115 +                                       continue;
 116 +
 117 +                               sym = syms + ELF_MIPS_R_SYM(rela[j]);
 118 +                               if (!is_init && sym->st_shndx != SHN_UNDEF)
 119 +                                       continue;
 120 +
 121 +                               ret += 4 * sizeof(int);
 122 +                       }
 123 +               }
 124 +       }
 125 +
 126 +       return ret;
 127 +}
 128 +
 129 +#ifndef MODULE_START
 130 +static void *alloc_phys(unsigned long size)
 131 +{
 132 +       unsigned order;
 133 +       struct page *page;
 134 +       struct page *p;
 135 +
 136 +       size = PAGE_ALIGN(size);
 137 +       order = get_order(size);
 138 +
 139 +       page = alloc_pages(GFP_KERNEL | __GFP_NORETRY | __GFP_NOWARN |
 140 +                       __GFP_THISNODE, order);
 141 +       if (!page)
 142 +               return NULL;
 143 +
 144 +       split_page(page, order);
 145 +
 146 +       /* mark all pages except for the last one */
 147 +       for (p = page; p + 1 < page + (size >> PAGE_SHIFT); ++p)
 148 +               set_bit(PG_owner_priv_1, &p->flags);
 149 +
 150 +       for (p = page + (size >> PAGE_SHIFT); p < page + (1 << order); ++p)
 151 +               __free_page(p);
 152 +
 153 +       return page_address(page);
 154 +}
 155 +#endif
 156 +
 157 +static void free_phys(void *ptr)
 158 +{
 159 +       struct page *page;
 160 +       bool free;
 161 +
 162 +       page = virt_to_page(ptr);
 163 +       do {
 164 +               free = test_and_clear_bit(PG_owner_priv_1, &page->flags);
 165 +               __free_page(page);
 166 +               page++;
 167 +       } while (free);
 168 +}
 169 +
 170 +
 171  void *module_alloc(unsigned long size)
 172  {
 173 +#ifdef MODULE_START
 174         return __vmalloc_node_range(size, 1, MODULE_START, MODULE_END,
 175                                 GFP_KERNEL, PAGE_KERNEL, 0, NUMA_NO_NODE,
 176                                 __builtin_return_address(0));
 177 +#else
 178 +       void *ptr;
 179 +
 180 +       if (size == 0)
 181 +               return NULL;
 182 +
 183 +       ptr = alloc_phys(size);
 184 +
 185 +       /* If we failed to allocate physically contiguous memory,
 186 +        * fall back to regular vmalloc. The module loader code will
 187 +        * create jump tables to handle long jumps */
 188 +       if (!ptr)
 189 +               return vmalloc(size);
 190 +
 191 +       return ptr;
 192 +#endif
 193  }
 194 +
 195 +static inline bool is_phys_addr(void *ptr)
 196 +{
 197 +#ifdef CONFIG_64BIT
 198 +       return (KSEGX((unsigned long)ptr) == CKSEG0);
 199 +#else
 200 +       return (KSEGX(ptr) == KSEG0);
 201  #endif
 202 +}
 203 +
 204 +/* Free memory returned from module_alloc */
 205 +void module_memfree(void *module_region)
 206 +{
 207 +       if (is_phys_addr(module_region))
 208 +               free_phys(module_region);
 209 +       else
 210 +               vfree(module_region);
 211 +}
 212 +
 213 +static void *__module_alloc(int size, bool phys)
 214 +{
 215 +       void *ptr;
 216 +
 217 +       if (phys)
 218 +               ptr = kmalloc(size, GFP_KERNEL);
 219 +       else
 220 +               ptr = vmalloc(size);
 221 +       return ptr;
 222 +}
 223 +
 224 +static void __module_free(void *ptr)
 225 +{
 226 +       if (is_phys_addr(ptr))
 227 +               kfree(ptr);
 228 +       else
 229 +               vfree(ptr);
 230 +}
 231 +
 232 +int module_frob_arch_sections(Elf_Ehdr *hdr, Elf_Shdr *sechdrs,
 233 +                             char *secstrings, struct module *mod)
 234 +{
 235 +       unsigned int symindex = 0;
 236 +       unsigned int core_size, init_size;
 237 +       int i;
 238 +
 239 +       mod->arch.phys_plt_offset = 0;
 240 +       mod->arch.virt_plt_offset = 0;
 241 +       mod->arch.phys_plt_tbl = NULL;
 242 +       mod->arch.virt_plt_tbl = NULL;
 243 +
 244 +       if (IS_ENABLED(CONFIG_64BIT))
 245 +               return 0;
 246 +
 247 +       for (i = 1; i < hdr->e_shnum; i++)
 248 +               if (sechdrs[i].sh_type == SHT_SYMTAB)
 249 +                       symindex = i;
 250 +
 251 +       core_size = get_plt_size(hdr, sechdrs, secstrings, symindex, false);
 252 +       init_size = get_plt_size(hdr, sechdrs, secstrings, symindex, true);
 253 +
 254 +       if ((core_size + init_size) == 0)
 255 +               return 0;
 256 +
 257 +       mod->arch.phys_plt_tbl = __module_alloc(core_size + init_size, 1);
 258 +       if (!mod->arch.phys_plt_tbl)
 259 +               return -ENOMEM;
 260 +
 261 +       mod->arch.virt_plt_tbl = __module_alloc(core_size + init_size, 0);
 262 +       if (!mod->arch.virt_plt_tbl) {
 263 +               __module_free(mod->arch.phys_plt_tbl);
 264 +               mod->arch.phys_plt_tbl = NULL;
 265 +               return -ENOMEM;
 266 +       }
 267 +
 268 +       return 0;
 269 +}
 270
 271  int apply_r_mips_none(struct module *me, u32 *location, Elf_Addr v)
 272  {
 273 @@ -65,8 +272,39 @@ static int apply_r_mips_32_rel(struct module *me, u32 *location, Elf_Addr v)
 274         return 0;
 275  }
 276
 277 +static Elf_Addr add_plt_entry_to(unsigned *plt_offset,
 278 +                                void *start, Elf_Addr v)
 279 +{
 280 +       unsigned *tramp = start + *plt_offset;
 281 +       *plt_offset += 4 * sizeof(int);
 282 +
 283 +       /* adjust carry for addiu */
 284 +       if (v & 0x00008000)
 285 +               v += 0x10000;
 286 +
 287 +       tramp[0] = 0x3c190000 | (v >> 16);      /* lui t9, hi16 */
 288 +       tramp[1] = 0x27390000 | (v & 0xffff);   /* addiu t9, t9, lo16 */
 289 +       tramp[2] = 0x03200008;                  /* jr t9 */
 290 +       tramp[3] = 0x00000000;                  /* nop */
 291 +
 292 +       return (Elf_Addr) tramp;
 293 +}
 294 +
 295 +static Elf_Addr add_plt_entry(struct module *me, void *location, Elf_Addr v)
 296 +{
 297 +       if (is_phys_addr(location))
 298 +               return add_plt_entry_to(&me->arch.phys_plt_offset,
 299 +                               me->arch.phys_plt_tbl, v);
 300 +       else
 301 +               return add_plt_entry_to(&me->arch.virt_plt_offset,
 302 +                               me->arch.virt_plt_tbl, v);
 303 +
 304 +}
 305 +
 306  static int apply_r_mips_26_rel(struct module *me, u32 *location, Elf_Addr v)
 307  {
 308 +       u32 ofs = *location & 0x03ffffff;
 309 +
 310         if (v % 4) {
 311                 pr_err("module %s: dangerous R_MIPS_26 REL relocation\n",
 312                        me->name);
 313 @@ -74,13 +312,17 @@ static int apply_r_mips_26_rel(struct module *me, u32 *location, Elf_Addr v)
 314         }
 315
 316         if ((v & 0xf0000000) != (((unsigned long)location + 4) & 0xf0000000)) {
 317 -               pr_err("module %s: relocation overflow\n",
 318 -                      me->name);
 319 -               return -ENOEXEC;
 320 +               v = add_plt_entry(me, location, v + (ofs << 2));
 321 +               if (!v) {
 322 +                       pr_err("module %s: relocation overflow\n",
 323 +                              me->name);
 324 +                       return -ENOEXEC;
 325 +               }
 326 +               ofs = 0;
 327         }
 328
 329         *location = (*location & ~0x03ffffff) |
 330 -                   ((*location + (v >> 2)) & 0x03ffffff);
 331 +                   ((ofs + (v >> 2)) & 0x03ffffff);
 332
 333         return 0;
 334  }
 335 @@ -349,9 +591,36 @@ int module_finalize(const Elf_Ehdr *hdr,
 336                 list_add(&me->arch.dbe_list, &dbe_list);
 337                 spin_unlock_irq(&dbe_lock);
 338         }
 339 +
 340 +       /* Get rid of the fixup trampoline if we're running the module
 341 +        * from physically mapped address space */
 342 +       if (me->arch.phys_plt_offset == 0) {
 343 +               __module_free(me->arch.phys_plt_tbl);
 344 +               me->arch.phys_plt_tbl = NULL;
 345 +       }
 346 +       if (me->arch.virt_plt_offset == 0) {
 347 +               __module_free(me->arch.virt_plt_tbl);
 348 +               me->arch.virt_plt_tbl = NULL;
 349 +       }
 350 +
 351         return 0;
 352  }
 353
 354 +void module_arch_freeing_init(struct module *mod)
 355 +{
 356 +       if (mod->state == MODULE_STATE_LIVE)
 357 +               return;
 358 +
 359 +       if (mod->arch.phys_plt_tbl) {
 360 +               __module_free(mod->arch.phys_plt_tbl);
 361 +               mod->arch.phys_plt_tbl = NULL;
 362 +       }
 363 +       if (mod->arch.virt_plt_tbl) {
 364 +               __module_free(mod->arch.virt_plt_tbl);
 365 +               mod->arch.virt_plt_tbl = NULL;
 366 +       }
 367 +}
 368 +
 369  void module_arch_cleanup(struct module *mod)
 370  {
 371         spin_lock_irq(&dbe_lock);
 372 --
 373 2.11.0
 374