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