Merge master.kernel.org:/pub/scm/linux/kernel/git/roland/infiniband
[linux-2.6] / drivers / mtd / mtdblock.c
1 /* 
2  * Direct MTD block device access
3  *
4  * $Id: mtdblock.c,v 1.66 2004/11/25 13:52:52 joern Exp $
5  *
6  * (C) 2000-2003 Nicolas Pitre <nico@cam.org>
7  * (C) 1999-2003 David Woodhouse <dwmw2@infradead.org>
8  */
9
10 #include <linux/config.h>
11 #include <linux/types.h>
12 #include <linux/module.h>
13 #include <linux/kernel.h>
14 #include <linux/fs.h>
15 #include <linux/init.h>
16 #include <linux/slab.h>
17 #include <linux/vmalloc.h>
18 #include <linux/mtd/mtd.h>
19 #include <linux/mtd/blktrans.h>
20
21 static struct mtdblk_dev {
22         struct mtd_info *mtd;
23         int count;
24         struct semaphore cache_sem;
25         unsigned char *cache_data;
26         unsigned long cache_offset;
27         unsigned int cache_size;
28         enum { STATE_EMPTY, STATE_CLEAN, STATE_DIRTY } cache_state;
29 } *mtdblks[MAX_MTD_DEVICES];
30
31 /*
32  * Cache stuff...
33  * 
34  * Since typical flash erasable sectors are much larger than what Linux's
35  * buffer cache can handle, we must implement read-modify-write on flash
36  * sectors for each block write requests.  To avoid over-erasing flash sectors
37  * and to speed things up, we locally cache a whole flash sector while it is
38  * being written to until a different sector is required.
39  */
40
41 static void erase_callback(struct erase_info *done)
42 {
43         wait_queue_head_t *wait_q = (wait_queue_head_t *)done->priv;
44         wake_up(wait_q);
45 }
46
47 static int erase_write (struct mtd_info *mtd, unsigned long pos, 
48                         int len, const char *buf)
49 {
50         struct erase_info erase;
51         DECLARE_WAITQUEUE(wait, current);
52         wait_queue_head_t wait_q;
53         size_t retlen;
54         int ret;
55
56         /*
57          * First, let's erase the flash block.
58          */
59
60         init_waitqueue_head(&wait_q);
61         erase.mtd = mtd;
62         erase.callback = erase_callback;
63         erase.addr = pos;
64         erase.len = len;
65         erase.priv = (u_long)&wait_q;
66
67         set_current_state(TASK_INTERRUPTIBLE);
68         add_wait_queue(&wait_q, &wait);
69
70         ret = MTD_ERASE(mtd, &erase);
71         if (ret) {
72                 set_current_state(TASK_RUNNING);
73                 remove_wait_queue(&wait_q, &wait);
74                 printk (KERN_WARNING "mtdblock: erase of region [0x%lx, 0x%x] "
75                                      "on \"%s\" failed\n",
76                         pos, len, mtd->name);
77                 return ret;
78         }
79
80         schedule();  /* Wait for erase to finish. */
81         remove_wait_queue(&wait_q, &wait);
82
83         /*
84          * Next, writhe data to flash.
85          */
86
87         ret = MTD_WRITE (mtd, pos, len, &retlen, buf);
88         if (ret)
89                 return ret;
90         if (retlen != len)
91                 return -EIO;
92         return 0;
93 }
94
95
96 static int write_cached_data (struct mtdblk_dev *mtdblk)
97 {
98         struct mtd_info *mtd = mtdblk->mtd;
99         int ret;
100
101         if (mtdblk->cache_state != STATE_DIRTY)
102                 return 0;
103
104         DEBUG(MTD_DEBUG_LEVEL2, "mtdblock: writing cached data for \"%s\" "
105                         "at 0x%lx, size 0x%x\n", mtd->name, 
106                         mtdblk->cache_offset, mtdblk->cache_size);
107         
108         ret = erase_write (mtd, mtdblk->cache_offset, 
109                            mtdblk->cache_size, mtdblk->cache_data);
110         if (ret)
111                 return ret;
112
113         /*
114          * Here we could argubly set the cache state to STATE_CLEAN.
115          * However this could lead to inconsistency since we will not 
116          * be notified if this content is altered on the flash by other 
117          * means.  Let's declare it empty and leave buffering tasks to
118          * the buffer cache instead.
119          */
120         mtdblk->cache_state = STATE_EMPTY;
121         return 0;
122 }
123
124
125 static int do_cached_write (struct mtdblk_dev *mtdblk, unsigned long pos, 
126                             int len, const char *buf)
127 {
128         struct mtd_info *mtd = mtdblk->mtd;
129         unsigned int sect_size = mtdblk->cache_size;
130         size_t retlen;
131         int ret;
132
133         DEBUG(MTD_DEBUG_LEVEL2, "mtdblock: write on \"%s\" at 0x%lx, size 0x%x\n",
134                 mtd->name, pos, len);
135         
136         if (!sect_size)
137                 return MTD_WRITE (mtd, pos, len, &retlen, buf);
138
139         while (len > 0) {
140                 unsigned long sect_start = (pos/sect_size)*sect_size;
141                 unsigned int offset = pos - sect_start;
142                 unsigned int size = sect_size - offset;
143                 if( size > len ) 
144                         size = len;
145
146                 if (size == sect_size) {
147                         /* 
148                          * We are covering a whole sector.  Thus there is no
149                          * need to bother with the cache while it may still be
150                          * useful for other partial writes.
151                          */
152                         ret = erase_write (mtd, pos, size, buf);
153                         if (ret)
154                                 return ret;
155                 } else {
156                         /* Partial sector: need to use the cache */
157
158                         if (mtdblk->cache_state == STATE_DIRTY &&
159                             mtdblk->cache_offset != sect_start) {
160                                 ret = write_cached_data(mtdblk);
161                                 if (ret) 
162                                         return ret;
163                         }
164
165                         if (mtdblk->cache_state == STATE_EMPTY ||
166                             mtdblk->cache_offset != sect_start) {
167                                 /* fill the cache with the current sector */
168                                 mtdblk->cache_state = STATE_EMPTY;
169                                 ret = MTD_READ(mtd, sect_start, sect_size, &retlen, mtdblk->cache_data);
170                                 if (ret)
171                                         return ret;
172                                 if (retlen != sect_size)
173                                         return -EIO;
174
175                                 mtdblk->cache_offset = sect_start;
176                                 mtdblk->cache_size = sect_size;
177                                 mtdblk->cache_state = STATE_CLEAN;
178                         }
179
180                         /* write data to our local cache */
181                         memcpy (mtdblk->cache_data + offset, buf, size);
182                         mtdblk->cache_state = STATE_DIRTY;
183                 }
184
185                 buf += size;
186                 pos += size;
187                 len -= size;
188         }
189
190         return 0;
191 }
192
193
194 static int do_cached_read (struct mtdblk_dev *mtdblk, unsigned long pos, 
195                            int len, char *buf)
196 {
197         struct mtd_info *mtd = mtdblk->mtd;
198         unsigned int sect_size = mtdblk->cache_size;
199         size_t retlen;
200         int ret;
201
202         DEBUG(MTD_DEBUG_LEVEL2, "mtdblock: read on \"%s\" at 0x%lx, size 0x%x\n", 
203                         mtd->name, pos, len);
204         
205         if (!sect_size)
206                 return MTD_READ (mtd, pos, len, &retlen, buf);
207
208         while (len > 0) {
209                 unsigned long sect_start = (pos/sect_size)*sect_size;
210                 unsigned int offset = pos - sect_start;
211                 unsigned int size = sect_size - offset;
212                 if (size > len) 
213                         size = len;
214
215                 /*
216                  * Check if the requested data is already cached
217                  * Read the requested amount of data from our internal cache if it
218                  * contains what we want, otherwise we read the data directly
219                  * from flash.
220                  */
221                 if (mtdblk->cache_state != STATE_EMPTY &&
222                     mtdblk->cache_offset == sect_start) {
223                         memcpy (buf, mtdblk->cache_data + offset, size);
224                 } else {
225                         ret = MTD_READ (mtd, pos, size, &retlen, buf);
226                         if (ret)
227                                 return ret;
228                         if (retlen != size)
229                                 return -EIO;
230                 }
231
232                 buf += size;
233                 pos += size;
234                 len -= size;
235         }
236
237         return 0;
238 }
239
240 static int mtdblock_readsect(struct mtd_blktrans_dev *dev,
241                               unsigned long block, char *buf)
242 {
243         struct mtdblk_dev *mtdblk = mtdblks[dev->devnum];
244         return do_cached_read(mtdblk, block<<9, 512, buf);
245 }
246
247 static int mtdblock_writesect(struct mtd_blktrans_dev *dev,
248                               unsigned long block, char *buf)
249 {
250         struct mtdblk_dev *mtdblk = mtdblks[dev->devnum];
251         if (unlikely(!mtdblk->cache_data && mtdblk->cache_size)) {
252                 mtdblk->cache_data = vmalloc(mtdblk->mtd->erasesize);
253                 if (!mtdblk->cache_data)
254                         return -EINTR;
255                 /* -EINTR is not really correct, but it is the best match
256                  * documented in man 2 write for all cases.  We could also
257                  * return -EAGAIN sometimes, but why bother?
258                  */
259         }
260         return do_cached_write(mtdblk, block<<9, 512, buf);
261 }
262
263 static int mtdblock_open(struct mtd_blktrans_dev *mbd)
264 {
265         struct mtdblk_dev *mtdblk;
266         struct mtd_info *mtd = mbd->mtd;
267         int dev = mbd->devnum;
268
269         DEBUG(MTD_DEBUG_LEVEL1,"mtdblock_open\n");
270         
271         if (mtdblks[dev]) {
272                 mtdblks[dev]->count++;
273                 return 0;
274         }
275         
276         /* OK, it's not open. Create cache info for it */
277         mtdblk = kmalloc(sizeof(struct mtdblk_dev), GFP_KERNEL);
278         if (!mtdblk)
279                 return -ENOMEM;
280
281         memset(mtdblk, 0, sizeof(*mtdblk));
282         mtdblk->count = 1;
283         mtdblk->mtd = mtd;
284
285         init_MUTEX (&mtdblk->cache_sem);
286         mtdblk->cache_state = STATE_EMPTY;
287         if ((mtdblk->mtd->flags & MTD_CAP_RAM) != MTD_CAP_RAM &&
288             mtdblk->mtd->erasesize) {
289                 mtdblk->cache_size = mtdblk->mtd->erasesize;
290                 mtdblk->cache_data = NULL;
291         }
292
293         mtdblks[dev] = mtdblk;
294         
295         DEBUG(MTD_DEBUG_LEVEL1, "ok\n");
296
297         return 0;
298 }
299
300 static int mtdblock_release(struct mtd_blktrans_dev *mbd)
301 {
302         int dev = mbd->devnum;
303         struct mtdblk_dev *mtdblk = mtdblks[dev];
304
305         DEBUG(MTD_DEBUG_LEVEL1, "mtdblock_release\n");
306
307         down(&mtdblk->cache_sem);
308         write_cached_data(mtdblk);
309         up(&mtdblk->cache_sem);
310
311         if (!--mtdblk->count) {
312                 /* It was the last usage. Free the device */
313                 mtdblks[dev] = NULL;
314                 if (mtdblk->mtd->sync)
315                         mtdblk->mtd->sync(mtdblk->mtd);
316                 vfree(mtdblk->cache_data);
317                 kfree(mtdblk);
318         }
319         DEBUG(MTD_DEBUG_LEVEL1, "ok\n");
320
321         return 0;
322 }  
323
324 static int mtdblock_flush(struct mtd_blktrans_dev *dev)
325 {
326         struct mtdblk_dev *mtdblk = mtdblks[dev->devnum];
327
328         down(&mtdblk->cache_sem);
329         write_cached_data(mtdblk);
330         up(&mtdblk->cache_sem);
331
332         if (mtdblk->mtd->sync)
333                 mtdblk->mtd->sync(mtdblk->mtd);
334         return 0;
335 }
336
337 static void mtdblock_add_mtd(struct mtd_blktrans_ops *tr, struct mtd_info *mtd)
338 {
339         struct mtd_blktrans_dev *dev = kmalloc(sizeof(*dev), GFP_KERNEL);
340
341         if (!dev)
342                 return;
343
344         memset(dev, 0, sizeof(*dev));
345
346         dev->mtd = mtd;
347         dev->devnum = mtd->index;
348         dev->blksize = 512;
349         dev->size = mtd->size >> 9;
350         dev->tr = tr;
351
352         if (!(mtd->flags & MTD_WRITEABLE))
353                 dev->readonly = 1;
354
355         add_mtd_blktrans_dev(dev);
356 }
357
358 static void mtdblock_remove_dev(struct mtd_blktrans_dev *dev)
359 {
360         del_mtd_blktrans_dev(dev);
361         kfree(dev);
362 }
363
364 static struct mtd_blktrans_ops mtdblock_tr = {
365         .name           = "mtdblock",
366         .major          = 31,
367         .part_bits      = 0,
368         .open           = mtdblock_open,
369         .flush          = mtdblock_flush,
370         .release        = mtdblock_release,
371         .readsect       = mtdblock_readsect,
372         .writesect      = mtdblock_writesect,
373         .add_mtd        = mtdblock_add_mtd,
374         .remove_dev     = mtdblock_remove_dev,
375         .owner          = THIS_MODULE,
376 };
377
378 static int __init init_mtdblock(void)
379 {
380         return register_mtd_blktrans(&mtdblock_tr);
381 }
382
383 static void __exit cleanup_mtdblock(void)
384 {
385         deregister_mtd_blktrans(&mtdblock_tr);
386 }
387
388 module_init(init_mtdblock);
389 module_exit(cleanup_mtdblock);
390
391
392 MODULE_LICENSE("GPL");
393 MODULE_AUTHOR("Nicolas Pitre <nico@cam.org> et al.");
394 MODULE_DESCRIPTION("Caching read/erase/writeback block device emulation access to MTD devices");