arch/um/kernel/physmem.c

   1 /*
   2  * Copyright (C) 2000 - 2003 Jeff Dike (jdike@addtoit.com)
   3  * Licensed under the GPL
   4  */
   5
   6 #include "linux/mm.h"
   7 #include "linux/rbtree.h"
   8 #include "linux/slab.h"
   9 #include "linux/vmalloc.h"
  10 #include "linux/bootmem.h"
  11 #include "linux/module.h"
  12 #include "asm/types.h"
  13 #include "asm/pgtable.h"
  14 #include "kern_util.h"
  15 #include "user_util.h"
  16 #include "mode_kern.h"
  17 #include "mem.h"
  18 #include "mem_user.h"
  19 #include "os.h"
  20 #include "kern.h"
  21 #include "init.h"
  22
  23 struct phys_desc {
  24         struct rb_node rb;
  25         int fd;
  26         __u64 offset;
  27         void *virt;
  28         unsigned long phys;
  29         struct list_head list;
  30 };
  31
  32 static struct rb_root phys_mappings = RB_ROOT;
  33
  34 static struct rb_node **find_rb(void *virt)
  35 {
  36         struct rb_node **n = &phys_mappings.rb_node;
  37         struct phys_desc *d;
  38
  39         while(*n != NULL){
  40                 d = rb_entry(*n, struct phys_desc, rb);
  41                 if(d->virt == virt)
  42                         return(n);
  43
  44                 if(d->virt > virt)
  45                         n = &(*n)->rb_left;
  46                 else
  47                         n = &(*n)->rb_right;
  48         }
  49
  50         return(n);
  51 }
  52
  53 static struct phys_desc *find_phys_mapping(void *virt)
  54 {
  55         struct rb_node **n = find_rb(virt);
  56
  57         if(*n == NULL)
  58                 return(NULL);
  59
  60         return(rb_entry(*n, struct phys_desc, rb));
  61 }
  62
  63 static void insert_phys_mapping(struct phys_desc *desc)
  64 {
  65         struct rb_node **n = find_rb(desc->virt);
  66
  67         if(*n != NULL)
  68                 panic("Physical remapping for %p already present",
  69                       desc->virt);
  70
  71         rb_link_node(&desc->rb, (*n)->rb_parent, n);
  72         rb_insert_color(&desc->rb, &phys_mappings);
  73 }
  74
  75 LIST_HEAD(descriptor_mappings);
  76
  77 struct desc_mapping {
  78         int fd;
  79         struct list_head list;
  80         struct list_head pages;
  81 };
  82
  83 static struct desc_mapping *find_mapping(int fd)
  84 {
  85         struct desc_mapping *desc;
  86         struct list_head *ele;
  87
  88         list_for_each(ele, &descriptor_mappings){
  89                 desc = list_entry(ele, struct desc_mapping, list);
  90                 if(desc->fd == fd)
  91                         return(desc);
  92         }
  93
  94         return(NULL);
  95 }
  96
  97 static struct desc_mapping *descriptor_mapping(int fd)
  98 {
  99         struct desc_mapping *desc;
 100
 101         desc = find_mapping(fd);
 102         if(desc != NULL)
 103                 return(desc);
 104
 105         desc = kmalloc(sizeof(*desc), GFP_ATOMIC);
 106         if(desc == NULL)
 107                 return(NULL);
 108
 109         *desc = ((struct desc_mapping)
 110                 { .fd =         fd,
 111                   .list =       LIST_HEAD_INIT(desc->list),
 112                   .pages =      LIST_HEAD_INIT(desc->pages) });
 113         list_add(&desc->list, &descriptor_mappings);
 114
 115         return(desc);
 116 }
 117
 118 int physmem_subst_mapping(void *virt, int fd, __u64 offset, int w)
 119 {
 120         struct desc_mapping *fd_maps;
 121         struct phys_desc *desc;
 122         unsigned long phys;
 123         int err;
 124
 125         fd_maps = descriptor_mapping(fd);
 126         if(fd_maps == NULL)
 127                 return(-ENOMEM);
 128
 129         phys = __pa(virt);
 130         desc = find_phys_mapping(virt);
 131         if(desc != NULL)
 132                 panic("Address 0x%p is already substituted\n", virt);
 133
 134         err = -ENOMEM;
 135         desc = kmalloc(sizeof(*desc), GFP_ATOMIC);
 136         if(desc == NULL)
 137                 goto out;
 138
 139         *desc = ((struct phys_desc)
 140                 { .fd =                 fd,
 141                   .offset =             offset,
 142                   .virt =               virt,
 143                   .phys =               __pa(virt),
 144                   .list =               LIST_HEAD_INIT(desc->list) });
 145         insert_phys_mapping(desc);
 146
 147         list_add(&desc->list, &fd_maps->pages);
 148
 149         virt = (void *) ((unsigned long) virt & PAGE_MASK);
 150         err = os_map_memory(virt, fd, offset, PAGE_SIZE, 1, w, 0);
 151         if(!err)
 152                 goto out;
 153
 154         rb_erase(&desc->rb, &phys_mappings);
 155         kfree(desc);
 156  out:
 157         return(err);
 158 }
 159
 160 static int physmem_fd = -1;
 161
 162 static void remove_mapping(struct phys_desc *desc)
 163 {
 164         void *virt = desc->virt;
 165         int err;
 166
 167         rb_erase(&desc->rb, &phys_mappings);
 168         list_del(&desc->list);
 169         kfree(desc);
 170
 171         err = os_map_memory(virt, physmem_fd, __pa(virt), PAGE_SIZE, 1, 1, 0);
 172         if(err)
 173                 panic("Failed to unmap block device page from physical memory, "
 174                       "errno = %d", -err);
 175 }
 176
 177 int physmem_remove_mapping(void *virt)
 178 {
 179         struct phys_desc *desc;
 180
 181         virt = (void *) ((unsigned long) virt & PAGE_MASK);
 182         desc = find_phys_mapping(virt);
 183         if(desc == NULL)
 184                 return(0);
 185
 186         remove_mapping(desc);
 187         return(1);
 188 }
 189
 190 void physmem_forget_descriptor(int fd)
 191 {
 192         struct desc_mapping *desc;
 193         struct phys_desc *page;
 194         struct list_head *ele, *next;
 195         __u64 offset;
 196         void *addr;
 197         int err;
 198
 199         desc = find_mapping(fd);
 200         if(desc == NULL)
 201                 return;
 202
 203         list_for_each_safe(ele, next, &desc->pages){
 204                 page = list_entry(ele, struct phys_desc, list);
 205                 offset = page->offset;
 206                 addr = page->virt;
 207                 remove_mapping(page);
 208                 err = os_seek_file(fd, offset);
 209                 if(err)
 210                         panic("physmem_forget_descriptor - failed to seek "
 211                               "to %lld in fd %d, error = %d\n",
 212                               offset, fd, -err);
 213                 err = os_read_file(fd, addr, PAGE_SIZE);
 214                 if(err < 0)
 215                         panic("physmem_forget_descriptor - failed to read "
 216                               "from fd %d to 0x%p, error = %d\n",
 217                               fd, addr, -err);
 218         }
 219
 220         list_del(&desc->list);
 221         kfree(desc);
 222 }
 223
 224 EXPORT_SYMBOL(physmem_forget_descriptor);
 225 EXPORT_SYMBOL(physmem_remove_mapping);
 226 EXPORT_SYMBOL(physmem_subst_mapping);
 227
 228 void arch_free_page(struct page *page, int order)
 229 {
 230         void *virt;
 231         int i;
 232
 233         for(i = 0; i < (1 << order); i++){
 234                 virt = __va(page_to_phys(page + i));
 235                 physmem_remove_mapping(virt);
 236         }
 237 }
 238
 239 int is_remapped(void *virt)
 240 {
 241         struct phys_desc *desc = find_phys_mapping(virt);
 242
 243         return(desc != NULL);
 244 }
 245
 246 /* Changed during early boot */
 247 unsigned long high_physmem;
 248
 249 extern unsigned long physmem_size;
 250
 251 void *to_virt(unsigned long phys)
 252 {
 253         return((void *) uml_physmem + phys);
 254 }
 255
 256 unsigned long to_phys(void *virt)
 257 {
 258         return(((unsigned long) virt) - uml_physmem);
 259 }
 260
 261 int init_maps(unsigned long physmem, unsigned long iomem, unsigned long highmem)
 262 {
 263         struct page *p, *map;
 264         unsigned long phys_len, phys_pages, highmem_len, highmem_pages;
 265         unsigned long iomem_len, iomem_pages, total_len, total_pages;
 266         int i;
 267
 268         phys_pages = physmem >> PAGE_SHIFT;
 269         phys_len = phys_pages * sizeof(struct page);
 270
 271         iomem_pages = iomem >> PAGE_SHIFT;
 272         iomem_len = iomem_pages * sizeof(struct page);
 273
 274         highmem_pages = highmem >> PAGE_SHIFT;
 275         highmem_len = highmem_pages * sizeof(struct page);
 276
 277         total_pages = phys_pages + iomem_pages + highmem_pages;
 278         total_len = phys_len + iomem_pages + highmem_len;
 279
 280         if(kmalloc_ok){
 281                 map = kmalloc(total_len, GFP_KERNEL);
 282                 if(map == NULL)
 283                         map = vmalloc(total_len);
 284         }
 285         else map = alloc_bootmem_low_pages(total_len);
 286
 287         if(map == NULL)
 288                 return(-ENOMEM);
 289
 290         for(i = 0; i < total_pages; i++){
 291                 p = &map[i];
 292                 set_page_count(p, 0);
 293                 SetPageReserved(p);
 294                 INIT_LIST_HEAD(&p->lru);
 295         }
 296
 297         max_mapnr = total_pages;
 298         return(0);
 299 }
 300
 301 struct page *phys_to_page(const unsigned long phys)
 302 {
 303         return(&mem_map[phys >> PAGE_SHIFT]);
 304 }
 305
 306 struct page *__virt_to_page(const unsigned long virt)
 307 {
 308         return(&mem_map[__pa(virt) >> PAGE_SHIFT]);
 309 }
 310
 311 phys_t page_to_phys(struct page *page)
 312 {
 313         return((page - mem_map) << PAGE_SHIFT);
 314 }
 315
 316 pte_t mk_pte(struct page *page, pgprot_t pgprot)
 317 {
 318         pte_t pte;
 319
 320         pte_set_val(pte, page_to_phys(page), pgprot);
 321         if(pte_present(pte))
 322                 pte_mknewprot(pte_mknewpage(pte));
 323         return(pte);
 324 }
 325
 326 /* Changed during early boot */
 327 static unsigned long kmem_top = 0;
 328
 329 unsigned long get_kmem_end(void)
 330 {
 331         if(kmem_top == 0)
 332                 kmem_top = CHOOSE_MODE(kmem_end_tt, kmem_end_skas);
 333         return(kmem_top);
 334 }
 335
 336 void map_memory(unsigned long virt, unsigned long phys, unsigned long len,
 337                 int r, int w, int x)
 338 {
 339         __u64 offset;
 340         int fd, err;
 341
 342         fd = phys_mapping(phys, &offset);
 343         err = os_map_memory((void *) virt, fd, offset, len, r, w, x);
 344         if(err) {
 345                 if(err == -ENOMEM)
 346                         printk("try increasing the host's "
 347                                "/proc/sys/vm/max_map_count to <physical "
 348                                "memory size>/4096\n");
 349                 panic("map_memory(0x%lx, %d, 0x%llx, %ld, %d, %d, %d) failed, "
 350                       "err = %d\n", virt, fd, offset, len, r, w, x, err);
 351         }
 352 }
 353
 354 #define PFN_UP(x) (((x) + PAGE_SIZE-1) >> PAGE_SHIFT)
 355
 356 extern int __syscall_stub_start, __binary_start;
 357
 358 void setup_physmem(unsigned long start, unsigned long reserve_end,
 359                    unsigned long len, unsigned long highmem)
 360 {
 361         unsigned long reserve = reserve_end - start;
 362         int pfn = PFN_UP(__pa(reserve_end));
 363         int delta = (len - reserve) >> PAGE_SHIFT;
 364         int err, offset, bootmap_size;
 365
 366         physmem_fd = create_mem_file(len + highmem);
 367
 368         offset = uml_reserved - uml_physmem;
 369         err = os_map_memory((void *) uml_reserved, physmem_fd, offset,
 370                             len - offset, 1, 1, 0);
 371         if(err < 0){
 372                 os_print_error(err, "Mapping memory");
 373                 exit(1);
 374         }
 375
 376         /* Special kludge - This page will be mapped in to userspace processes
 377          * from physmem_fd, so it needs to be written out there.
 378          */
 379         os_seek_file(physmem_fd, __pa(&__syscall_stub_start));
 380         os_write_file(physmem_fd, &__syscall_stub_start, PAGE_SIZE);
 381
 382         bootmap_size = init_bootmem(pfn, pfn + delta);
 383         free_bootmem(__pa(reserve_end) + bootmap_size,
 384                      len - bootmap_size - reserve);
 385 }
 386
 387 int phys_mapping(unsigned long phys, __u64 *offset_out)
 388 {
 389         struct phys_desc *desc = find_phys_mapping(__va(phys & PAGE_MASK));
 390         int fd = -1;
 391
 392         if(desc != NULL){
 393                 fd = desc->fd;
 394                 *offset_out = desc->offset;
 395         }
 396         else if(phys < physmem_size){
 397                 fd = physmem_fd;
 398                 *offset_out = phys;
 399         }
 400         else if(phys < __pa(end_iomem)){
 401                 struct iomem_region *region = iomem_regions;
 402
 403                 while(region != NULL){
 404                         if((phys >= region->phys) &&
 405                            (phys < region->phys + region->size)){
 406                                 fd = region->fd;
 407                                 *offset_out = phys - region->phys;
 408                                 break;
 409                         }
 410                         region = region->next;
 411                 }
 412         }
 413         else if(phys < __pa(end_iomem) + highmem){
 414                 fd = physmem_fd;
 415                 *offset_out = phys - iomem_size;
 416         }
 417
 418         return(fd);
 419 }
 420
 421 static int __init uml_mem_setup(char *line, int *add)
 422 {
 423         char *retptr;
 424         physmem_size = memparse(line,&retptr);
 425         return 0;
 426 }
 427 __uml_setup("mem=", uml_mem_setup,
 428 "mem=<Amount of desired ram>\n"
 429 "    This controls how much \"physical\" memory the kernel allocates\n"
 430 "    for the system. The size is specified as a number followed by\n"
 431 "    one of 'k', 'K', 'm', 'M', which have the obvious meanings.\n"
 432 "    This is not related to the amount of memory in the host.  It can\n"
 433 "    be more, and the excess, if it's ever used, will just be swapped out.\n"
 434 "       Example: mem=64M\n\n"
 435 );
 436
 437 unsigned long find_iomem(char *driver, unsigned long *len_out)
 438 {
 439         struct iomem_region *region = iomem_regions;
 440
 441         while(region != NULL){
 442                 if(!strcmp(region->driver, driver)){
 443                         *len_out = region->size;
 444                         return(region->virt);
 445                 }
 446         }
 447
 448         return(0);
 449 }
 450
 451 int setup_iomem(void)
 452 {
 453         struct iomem_region *region = iomem_regions;
 454         unsigned long iomem_start = high_physmem + PAGE_SIZE;
 455         int err;
 456
 457         while(region != NULL){
 458                 err = os_map_memory((void *) iomem_start, region->fd, 0,
 459                                     region->size, 1, 1, 0);
 460                 if(err)
 461                         printk("Mapping iomem region for driver '%s' failed, "
 462                                "errno = %d\n", region->driver, -err);
 463                 else {
 464                         region->virt = iomem_start;
 465                         region->phys = __pa(region->virt);
 466                 }
 467
 468                 iomem_start += region->size + PAGE_SIZE;
 469                 region = region->next;
 470         }
 471
 472         return(0);
 473 }
 474
 475 __initcall(setup_iomem);
 476
 477 /*
 478  * Overrides for Emacs so that we follow Linus's tabbing style.
 479  * Emacs will notice this stuff at the end of the file and automatically
 480  * adjust the settings for this buffer only.  This must remain at the end
 481  * of the file.
 482  * ---------------------------------------------------------------------------
 483  * Local variables:
 484  * c-file-style: "linux"
 485  * End:
 486  */