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