Merge git://git.kernel.org/pub/scm/linux/kernel/git/steve/gfs2-2.6-fixes
[linux-2.6] / 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 "linux/pfn.h"
13 #include "asm/types.h"
14 #include "asm/pgtable.h"
15 #include "kern_util.h"
16 #include "user_util.h"
17 #include "mode_kern.h"
18 #include "mem.h"
19 #include "mem_user.h"
20 #include "os.h"
21 #include "kern.h"
22 #include "init.h"
23
24 struct phys_desc {
25         struct rb_node rb;
26         int fd;
27         __u64 offset;
28         void *virt;
29         unsigned long phys;
30         struct list_head list;
31 };
32
33 static struct rb_root phys_mappings = RB_ROOT;
34
35 static struct rb_node **find_rb(void *virt)
36 {
37         struct rb_node **n = &phys_mappings.rb_node;
38         struct phys_desc *d;
39
40         while(*n != NULL){
41                 d = rb_entry(*n, struct phys_desc, rb);
42                 if(d->virt == virt)
43                         return(n);
44
45                 if(d->virt > virt)
46                         n = &(*n)->rb_left;
47                 else
48                         n = &(*n)->rb_right;
49         }
50
51         return(n);
52 }
53
54 static struct phys_desc *find_phys_mapping(void *virt)
55 {
56         struct rb_node **n = find_rb(virt);
57
58         if(*n == NULL)
59                 return(NULL);
60
61         return(rb_entry(*n, struct phys_desc, rb));
62 }
63
64 static void insert_phys_mapping(struct phys_desc *desc)
65 {
66         struct rb_node **n = find_rb(desc->virt);
67
68         if(*n != NULL)
69                 panic("Physical remapping for %p already present",
70                       desc->virt);
71
72         rb_link_node(&desc->rb, rb_parent(*n), n);
73         rb_insert_color(&desc->rb, &phys_mappings);
74 }
75
76 LIST_HEAD(descriptor_mappings);
77
78 struct desc_mapping {
79         int fd;
80         struct list_head list;
81         struct list_head pages;
82 };
83
84 static struct desc_mapping *find_mapping(int fd)
85 {
86         struct desc_mapping *desc;
87         struct list_head *ele;
88
89         list_for_each(ele, &descriptor_mappings){
90                 desc = list_entry(ele, struct desc_mapping, list);
91                 if(desc->fd == fd)
92                         return(desc);
93         }
94
95         return(NULL);
96 }
97
98 static struct desc_mapping *descriptor_mapping(int fd)
99 {
100         struct desc_mapping *desc;
101
102         desc = find_mapping(fd);
103         if(desc != NULL)
104                 return(desc);
105
106         desc = kmalloc(sizeof(*desc), GFP_ATOMIC);
107         if(desc == NULL)
108                 return(NULL);
109
110         *desc = ((struct desc_mapping)
111                 { .fd =         fd,
112                   .list =       LIST_HEAD_INIT(desc->list),
113                   .pages =      LIST_HEAD_INIT(desc->pages) });
114         list_add(&desc->list, &descriptor_mappings);
115
116         return(desc);
117 }
118
119 int physmem_subst_mapping(void *virt, int fd, __u64 offset, int w)
120 {
121         struct desc_mapping *fd_maps;
122         struct phys_desc *desc;
123         unsigned long phys;
124         int err;
125
126         fd_maps = descriptor_mapping(fd);
127         if(fd_maps == NULL)
128                 return(-ENOMEM);
129
130         phys = __pa(virt);
131         desc = find_phys_mapping(virt);
132         if(desc != NULL)
133                 panic("Address 0x%p is already substituted\n", virt);
134
135         err = -ENOMEM;
136         desc = kmalloc(sizeof(*desc), GFP_ATOMIC);
137         if(desc == NULL)
138                 goto out;
139
140         *desc = ((struct phys_desc)
141                 { .fd =                 fd,
142                   .offset =             offset,
143                   .virt =               virt,
144                   .phys =               __pa(virt),
145                   .list =               LIST_HEAD_INIT(desc->list) });
146         insert_phys_mapping(desc);
147
148         list_add(&desc->list, &fd_maps->pages);
149
150         virt = (void *) ((unsigned long) virt & PAGE_MASK);
151         err = os_map_memory(virt, fd, offset, PAGE_SIZE, 1, w, 0);
152         if(!err)
153                 goto out;
154
155         rb_erase(&desc->rb, &phys_mappings);
156         kfree(desc);
157  out:
158         return(err);
159 }
160
161 static int physmem_fd = -1;
162
163 static void remove_mapping(struct phys_desc *desc)
164 {
165         void *virt = desc->virt;
166         int err;
167
168         rb_erase(&desc->rb, &phys_mappings);
169         list_del(&desc->list);
170         kfree(desc);
171
172         err = os_map_memory(virt, physmem_fd, __pa(virt), PAGE_SIZE, 1, 1, 0);
173         if(err)
174                 panic("Failed to unmap block device page from physical memory, "
175                       "errno = %d", -err);
176 }
177
178 int physmem_remove_mapping(void *virt)
179 {
180         struct phys_desc *desc;
181
182         virt = (void *) ((unsigned long) virt & PAGE_MASK);
183         desc = find_phys_mapping(virt);
184         if(desc == NULL)
185                 return(0);
186
187         remove_mapping(desc);
188         return(1);
189 }
190
191 void physmem_forget_descriptor(int fd)
192 {
193         struct desc_mapping *desc;
194         struct phys_desc *page;
195         struct list_head *ele, *next;
196         __u64 offset;
197         void *addr;
198         int err;
199
200         desc = find_mapping(fd);
201         if(desc == NULL)
202                 return;
203
204         list_for_each_safe(ele, next, &desc->pages){
205                 page = list_entry(ele, struct phys_desc, list);
206                 offset = page->offset;
207                 addr = page->virt;
208                 remove_mapping(page);
209                 err = os_seek_file(fd, offset);
210                 if(err)
211                         panic("physmem_forget_descriptor - failed to seek "
212                               "to %lld in fd %d, error = %d\n",
213                               offset, fd, -err);
214                 err = os_read_file(fd, addr, PAGE_SIZE);
215                 if(err < 0)
216                         panic("physmem_forget_descriptor - failed to read "
217                               "from fd %d to 0x%p, error = %d\n",
218                               fd, addr, -err);
219         }
220
221         list_del(&desc->list);
222         kfree(desc);
223 }
224
225 EXPORT_SYMBOL(physmem_forget_descriptor);
226 EXPORT_SYMBOL(physmem_remove_mapping);
227 EXPORT_SYMBOL(physmem_subst_mapping);
228
229 void arch_free_page(struct page *page, int order)
230 {
231         void *virt;
232         int i;
233
234         for(i = 0; i < (1 << order); i++){
235                 virt = __va(page_to_phys(page + i));
236                 physmem_remove_mapping(virt);
237         }
238 }
239
240 int is_remapped(void *virt)
241 {
242         struct phys_desc *desc = find_phys_mapping(virt);
243
244         return(desc != NULL);
245 }
246
247 /* Changed during early boot */
248 unsigned long high_physmem;
249
250 extern unsigned long long physmem_size;
251
252 int init_maps(unsigned long physmem, unsigned long iomem, unsigned long highmem)
253 {
254         struct page *p, *map;
255         unsigned long phys_len, phys_pages, highmem_len, highmem_pages;
256         unsigned long iomem_len, iomem_pages, total_len, total_pages;
257         int i;
258
259         phys_pages = physmem >> PAGE_SHIFT;
260         phys_len = phys_pages * sizeof(struct page);
261
262         iomem_pages = iomem >> PAGE_SHIFT;
263         iomem_len = iomem_pages * sizeof(struct page);
264
265         highmem_pages = highmem >> PAGE_SHIFT;
266         highmem_len = highmem_pages * sizeof(struct page);
267
268         total_pages = phys_pages + iomem_pages + highmem_pages;
269         total_len = phys_len + iomem_len + highmem_len;
270
271         if(kmalloc_ok){
272                 map = kmalloc(total_len, GFP_KERNEL);
273                 if(map == NULL)
274                         map = vmalloc(total_len);
275         }
276         else map = alloc_bootmem_low_pages(total_len);
277
278         if(map == NULL)
279                 return(-ENOMEM);
280
281         for(i = 0; i < total_pages; i++){
282                 p = &map[i];
283                 memset(p, 0, sizeof(struct page));
284                 SetPageReserved(p);
285                 INIT_LIST_HEAD(&p->lru);
286         }
287
288         max_mapnr = total_pages;
289         return(0);
290 }
291
292 /* Changed during early boot */
293 static unsigned long kmem_top = 0;
294
295 unsigned long get_kmem_end(void)
296 {
297         if(kmem_top == 0)
298                 kmem_top = CHOOSE_MODE(kmem_end_tt, kmem_end_skas);
299         return(kmem_top);
300 }
301
302 void map_memory(unsigned long virt, unsigned long phys, unsigned long len,
303                 int r, int w, int x)
304 {
305         __u64 offset;
306         int fd, err;
307
308         fd = phys_mapping(phys, &offset);
309         err = os_map_memory((void *) virt, fd, offset, len, r, w, x);
310         if(err) {
311                 if(err == -ENOMEM)
312                         printk("try increasing the host's "
313                                "/proc/sys/vm/max_map_count to <physical "
314                                "memory size>/4096\n");
315                 panic("map_memory(0x%lx, %d, 0x%llx, %ld, %d, %d, %d) failed, "
316                       "err = %d\n", virt, fd, offset, len, r, w, x, err);
317         }
318 }
319
320 extern int __syscall_stub_start;
321
322 void setup_physmem(unsigned long start, unsigned long reserve_end,
323                    unsigned long len, unsigned long long highmem)
324 {
325         unsigned long reserve = reserve_end - start;
326         int pfn = PFN_UP(__pa(reserve_end));
327         int delta = (len - reserve) >> PAGE_SHIFT;
328         int err, offset, bootmap_size;
329
330         physmem_fd = create_mem_file(len + highmem);
331
332         offset = uml_reserved - uml_physmem;
333         err = os_map_memory((void *) uml_reserved, physmem_fd, offset,
334                             len - offset, 1, 1, 0);
335         if(err < 0){
336                 os_print_error(err, "Mapping memory");
337                 exit(1);
338         }
339
340         /* Special kludge - This page will be mapped in to userspace processes
341          * from physmem_fd, so it needs to be written out there.
342          */
343         os_seek_file(physmem_fd, __pa(&__syscall_stub_start));
344         os_write_file(physmem_fd, &__syscall_stub_start, PAGE_SIZE);
345
346         bootmap_size = init_bootmem(pfn, pfn + delta);
347         free_bootmem(__pa(reserve_end) + bootmap_size,
348                      len - bootmap_size - reserve);
349 }
350
351 int phys_mapping(unsigned long phys, __u64 *offset_out)
352 {
353         struct phys_desc *desc = find_phys_mapping(__va(phys & PAGE_MASK));
354         int fd = -1;
355
356         if(desc != NULL){
357                 fd = desc->fd;
358                 *offset_out = desc->offset;
359         }
360         else if(phys < physmem_size){
361                 fd = physmem_fd;
362                 *offset_out = phys;
363         }
364         else if(phys < __pa(end_iomem)){
365                 struct iomem_region *region = iomem_regions;
366
367                 while(region != NULL){
368                         if((phys >= region->phys) &&
369                            (phys < region->phys + region->size)){
370                                 fd = region->fd;
371                                 *offset_out = phys - region->phys;
372                                 break;
373                         }
374                         region = region->next;
375                 }
376         }
377         else if(phys < __pa(end_iomem) + highmem){
378                 fd = physmem_fd;
379                 *offset_out = phys - iomem_size;
380         }
381
382         return(fd);
383 }
384
385 static int __init uml_mem_setup(char *line, int *add)
386 {
387         char *retptr;
388         physmem_size = memparse(line,&retptr);
389         return 0;
390 }
391 __uml_setup("mem=", uml_mem_setup,
392 "mem=<Amount of desired ram>\n"
393 "    This controls how much \"physical\" memory the kernel allocates\n"
394 "    for the system. The size is specified as a number followed by\n"
395 "    one of 'k', 'K', 'm', 'M', which have the obvious meanings.\n"
396 "    This is not related to the amount of memory in the host.  It can\n"
397 "    be more, and the excess, if it's ever used, will just be swapped out.\n"
398 "       Example: mem=64M\n\n"
399 );
400
401 unsigned long find_iomem(char *driver, unsigned long *len_out)
402 {
403         struct iomem_region *region = iomem_regions;
404
405         while(region != NULL){
406                 if(!strcmp(region->driver, driver)){
407                         *len_out = region->size;
408                         return(region->virt);
409                 }
410
411                 region = region->next;
412         }
413
414         return(0);
415 }
416
417 int setup_iomem(void)
418 {
419         struct iomem_region *region = iomem_regions;
420         unsigned long iomem_start = high_physmem + PAGE_SIZE;
421         int err;
422
423         while(region != NULL){
424                 err = os_map_memory((void *) iomem_start, region->fd, 0,
425                                     region->size, 1, 1, 0);
426                 if(err)
427                         printk("Mapping iomem region for driver '%s' failed, "
428                                "errno = %d\n", region->driver, -err);
429                 else {
430                         region->virt = iomem_start;
431                         region->phys = __pa(region->virt);
432                 }
433
434                 iomem_start += region->size + PAGE_SIZE;
435                 region = region->next;
436         }
437
438         return(0);
439 }
440
441 __initcall(setup_iomem);
442
443 /*
444  * Overrides for Emacs so that we follow Linus's tabbing style.
445  * Emacs will notice this stuff at the end of the file and automatically
446  * adjust the settings for this buffer only.  This must remain at the end
447  * of the file.
448  * ---------------------------------------------------------------------------
449  * Local variables:
450  * c-file-style: "linux"
451  * End:
452  */