block: make blkdev_get/put() handle exclusive access
[linux-2.6.git] / fs / logfs / dev_bdev.c
1 /*
2  * fs/logfs/dev_bdev.c  - Device access methods for block devices
3  *
4  * As should be obvious for Linux kernel code, license is GPLv2
5  *
6  * Copyright (c) 2005-2008 Joern Engel <joern@logfs.org>
7  */
8 #include "logfs.h"
9 #include <linux/bio.h>
10 #include <linux/blkdev.h>
11 #include <linux/buffer_head.h>
12 #include <linux/gfp.h>
13
14 #define PAGE_OFS(ofs) ((ofs) & (PAGE_SIZE-1))
15
16 static void request_complete(struct bio *bio, int err)
17 {
18         complete((struct completion *)bio->bi_private);
19 }
20
21 static int sync_request(struct page *page, struct block_device *bdev, int rw)
22 {
23         struct bio bio;
24         struct bio_vec bio_vec;
25         struct completion complete;
26
27         bio_init(&bio);
28         bio.bi_io_vec = &bio_vec;
29         bio_vec.bv_page = page;
30         bio_vec.bv_len = PAGE_SIZE;
31         bio_vec.bv_offset = 0;
32         bio.bi_vcnt = 1;
33         bio.bi_idx = 0;
34         bio.bi_size = PAGE_SIZE;
35         bio.bi_bdev = bdev;
36         bio.bi_sector = page->index * (PAGE_SIZE >> 9);
37         init_completion(&complete);
38         bio.bi_private = &complete;
39         bio.bi_end_io = request_complete;
40
41         submit_bio(rw, &bio);
42         generic_unplug_device(bdev_get_queue(bdev));
43         wait_for_completion(&complete);
44         return test_bit(BIO_UPTODATE, &bio.bi_flags) ? 0 : -EIO;
45 }
46
47 static int bdev_readpage(void *_sb, struct page *page)
48 {
49         struct super_block *sb = _sb;
50         struct block_device *bdev = logfs_super(sb)->s_bdev;
51         int err;
52
53         err = sync_request(page, bdev, READ);
54         if (err) {
55                 ClearPageUptodate(page);
56                 SetPageError(page);
57         } else {
58                 SetPageUptodate(page);
59                 ClearPageError(page);
60         }
61         unlock_page(page);
62         return err;
63 }
64
65 static DECLARE_WAIT_QUEUE_HEAD(wq);
66
67 static void writeseg_end_io(struct bio *bio, int err)
68 {
69         const int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
70         struct bio_vec *bvec = bio->bi_io_vec + bio->bi_vcnt - 1;
71         struct super_block *sb = bio->bi_private;
72         struct logfs_super *super = logfs_super(sb);
73         struct page *page;
74
75         BUG_ON(!uptodate); /* FIXME: Retry io or write elsewhere */
76         BUG_ON(err);
77         BUG_ON(bio->bi_vcnt == 0);
78         do {
79                 page = bvec->bv_page;
80                 if (--bvec >= bio->bi_io_vec)
81                         prefetchw(&bvec->bv_page->flags);
82
83                 end_page_writeback(page);
84                 page_cache_release(page);
85         } while (bvec >= bio->bi_io_vec);
86         bio_put(bio);
87         if (atomic_dec_and_test(&super->s_pending_writes))
88                 wake_up(&wq);
89 }
90
91 static int __bdev_writeseg(struct super_block *sb, u64 ofs, pgoff_t index,
92                 size_t nr_pages)
93 {
94         struct logfs_super *super = logfs_super(sb);
95         struct address_space *mapping = super->s_mapping_inode->i_mapping;
96         struct bio *bio;
97         struct page *page;
98         struct request_queue *q = bdev_get_queue(sb->s_bdev);
99         unsigned int max_pages = queue_max_hw_sectors(q) >> (PAGE_SHIFT - 9);
100         int i;
101
102         if (max_pages > BIO_MAX_PAGES)
103                 max_pages = BIO_MAX_PAGES;
104         bio = bio_alloc(GFP_NOFS, max_pages);
105         BUG_ON(!bio);
106
107         for (i = 0; i < nr_pages; i++) {
108                 if (i >= max_pages) {
109                         /* Block layer cannot split bios :( */
110                         bio->bi_vcnt = i;
111                         bio->bi_idx = 0;
112                         bio->bi_size = i * PAGE_SIZE;
113                         bio->bi_bdev = super->s_bdev;
114                         bio->bi_sector = ofs >> 9;
115                         bio->bi_private = sb;
116                         bio->bi_end_io = writeseg_end_io;
117                         atomic_inc(&super->s_pending_writes);
118                         submit_bio(WRITE, bio);
119
120                         ofs += i * PAGE_SIZE;
121                         index += i;
122                         nr_pages -= i;
123                         i = 0;
124
125                         bio = bio_alloc(GFP_NOFS, max_pages);
126                         BUG_ON(!bio);
127                 }
128                 page = find_lock_page(mapping, index + i);
129                 BUG_ON(!page);
130                 bio->bi_io_vec[i].bv_page = page;
131                 bio->bi_io_vec[i].bv_len = PAGE_SIZE;
132                 bio->bi_io_vec[i].bv_offset = 0;
133
134                 BUG_ON(PageWriteback(page));
135                 set_page_writeback(page);
136                 unlock_page(page);
137         }
138         bio->bi_vcnt = nr_pages;
139         bio->bi_idx = 0;
140         bio->bi_size = nr_pages * PAGE_SIZE;
141         bio->bi_bdev = super->s_bdev;
142         bio->bi_sector = ofs >> 9;
143         bio->bi_private = sb;
144         bio->bi_end_io = writeseg_end_io;
145         atomic_inc(&super->s_pending_writes);
146         submit_bio(WRITE, bio);
147         return 0;
148 }
149
150 static void bdev_writeseg(struct super_block *sb, u64 ofs, size_t len)
151 {
152         struct logfs_super *super = logfs_super(sb);
153         int head;
154
155         BUG_ON(super->s_flags & LOGFS_SB_FLAG_RO);
156
157         if (len == 0) {
158                 /* This can happen when the object fit perfectly into a
159                  * segment, the segment gets written per sync and subsequently
160                  * closed.
161                  */
162                 return;
163         }
164         head = ofs & (PAGE_SIZE - 1);
165         if (head) {
166                 ofs -= head;
167                 len += head;
168         }
169         len = PAGE_ALIGN(len);
170         __bdev_writeseg(sb, ofs, ofs >> PAGE_SHIFT, len >> PAGE_SHIFT);
171         generic_unplug_device(bdev_get_queue(logfs_super(sb)->s_bdev));
172 }
173
174
175 static void erase_end_io(struct bio *bio, int err) 
176
177         const int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags); 
178         struct super_block *sb = bio->bi_private; 
179         struct logfs_super *super = logfs_super(sb); 
180
181         BUG_ON(!uptodate); /* FIXME: Retry io or write elsewhere */ 
182         BUG_ON(err); 
183         BUG_ON(bio->bi_vcnt == 0); 
184         bio_put(bio); 
185         if (atomic_dec_and_test(&super->s_pending_writes))
186                 wake_up(&wq); 
187
188
189 static int do_erase(struct super_block *sb, u64 ofs, pgoff_t index,
190                 size_t nr_pages)
191 {
192         struct logfs_super *super = logfs_super(sb);
193         struct bio *bio;
194         struct request_queue *q = bdev_get_queue(sb->s_bdev);
195         unsigned int max_pages = queue_max_hw_sectors(q) >> (PAGE_SHIFT - 9);
196         int i;
197
198         if (max_pages > BIO_MAX_PAGES)
199                 max_pages = BIO_MAX_PAGES;
200         bio = bio_alloc(GFP_NOFS, max_pages);
201         BUG_ON(!bio);
202
203         for (i = 0; i < nr_pages; i++) {
204                 if (i >= max_pages) {
205                         /* Block layer cannot split bios :( */
206                         bio->bi_vcnt = i;
207                         bio->bi_idx = 0;
208                         bio->bi_size = i * PAGE_SIZE;
209                         bio->bi_bdev = super->s_bdev;
210                         bio->bi_sector = ofs >> 9;
211                         bio->bi_private = sb;
212                         bio->bi_end_io = erase_end_io;
213                         atomic_inc(&super->s_pending_writes);
214                         submit_bio(WRITE, bio);
215
216                         ofs += i * PAGE_SIZE;
217                         index += i;
218                         nr_pages -= i;
219                         i = 0;
220
221                         bio = bio_alloc(GFP_NOFS, max_pages);
222                         BUG_ON(!bio);
223                 }
224                 bio->bi_io_vec[i].bv_page = super->s_erase_page;
225                 bio->bi_io_vec[i].bv_len = PAGE_SIZE;
226                 bio->bi_io_vec[i].bv_offset = 0;
227         }
228         bio->bi_vcnt = nr_pages;
229         bio->bi_idx = 0;
230         bio->bi_size = nr_pages * PAGE_SIZE;
231         bio->bi_bdev = super->s_bdev;
232         bio->bi_sector = ofs >> 9;
233         bio->bi_private = sb;
234         bio->bi_end_io = erase_end_io;
235         atomic_inc(&super->s_pending_writes);
236         submit_bio(WRITE, bio);
237         return 0;
238 }
239
240 static int bdev_erase(struct super_block *sb, loff_t to, size_t len,
241                 int ensure_write)
242 {
243         struct logfs_super *super = logfs_super(sb);
244
245         BUG_ON(to & (PAGE_SIZE - 1));
246         BUG_ON(len & (PAGE_SIZE - 1));
247
248         if (super->s_flags & LOGFS_SB_FLAG_RO)
249                 return -EROFS;
250
251         if (ensure_write) {
252                 /*
253                  * Object store doesn't care whether erases happen or not.
254                  * But for the journal they are required.  Otherwise a scan
255                  * can find an old commit entry and assume it is the current
256                  * one, travelling back in time.
257                  */
258                 do_erase(sb, to, to >> PAGE_SHIFT, len >> PAGE_SHIFT);
259         }
260
261         return 0;
262 }
263
264 static void bdev_sync(struct super_block *sb)
265 {
266         struct logfs_super *super = logfs_super(sb);
267
268         wait_event(wq, atomic_read(&super->s_pending_writes) == 0);
269 }
270
271 static struct page *bdev_find_first_sb(struct super_block *sb, u64 *ofs)
272 {
273         struct logfs_super *super = logfs_super(sb);
274         struct address_space *mapping = super->s_mapping_inode->i_mapping;
275         filler_t *filler = bdev_readpage;
276
277         *ofs = 0;
278         return read_cache_page(mapping, 0, filler, sb);
279 }
280
281 static struct page *bdev_find_last_sb(struct super_block *sb, u64 *ofs)
282 {
283         struct logfs_super *super = logfs_super(sb);
284         struct address_space *mapping = super->s_mapping_inode->i_mapping;
285         filler_t *filler = bdev_readpage;
286         u64 pos = (super->s_bdev->bd_inode->i_size & ~0xfffULL) - 0x1000;
287         pgoff_t index = pos >> PAGE_SHIFT;
288
289         *ofs = pos;
290         return read_cache_page(mapping, index, filler, sb);
291 }
292
293 static int bdev_write_sb(struct super_block *sb, struct page *page)
294 {
295         struct block_device *bdev = logfs_super(sb)->s_bdev;
296
297         /* Nothing special to do for block devices. */
298         return sync_request(page, bdev, WRITE);
299 }
300
301 static void bdev_put_device(struct logfs_super *s)
302 {
303         blkdev_put(s->s_bdev, FMODE_READ|FMODE_WRITE|FMODE_EXCL);
304 }
305
306 static int bdev_can_write_buf(struct super_block *sb, u64 ofs)
307 {
308         return 0;
309 }
310
311 static const struct logfs_device_ops bd_devops = {
312         .find_first_sb  = bdev_find_first_sb,
313         .find_last_sb   = bdev_find_last_sb,
314         .write_sb       = bdev_write_sb,
315         .readpage       = bdev_readpage,
316         .writeseg       = bdev_writeseg,
317         .erase          = bdev_erase,
318         .can_write_buf  = bdev_can_write_buf,
319         .sync           = bdev_sync,
320         .put_device     = bdev_put_device,
321 };
322
323 int logfs_get_sb_bdev(struct logfs_super *p, struct file_system_type *type,
324                 const char *devname)
325 {
326         struct block_device *bdev;
327
328         bdev = open_bdev_exclusive(devname, FMODE_READ|FMODE_WRITE, type);
329         if (IS_ERR(bdev))
330                 return PTR_ERR(bdev);
331
332         if (MAJOR(bdev->bd_dev) == MTD_BLOCK_MAJOR) {
333                 int mtdnr = MINOR(bdev->bd_dev);
334                 blkdev_put(bdev, FMODE_READ|FMODE_WRITE|FMODE_EXCL);
335                 return logfs_get_sb_mtd(p, mtdnr);
336         }
337
338         p->s_bdev = bdev;
339         p->s_mtd = NULL;
340         p->s_devops = &bd_devops;
341         return 0;
342 }