fsfilt_ext3.c

来自「lustre 1.6.5 source code」· C语言 代码 · 共 1,853 行 · 第 1/5 页

 * any of the matching commit_write operations, so even if we split
 * up to use "smaller" transactions none of them could complete until
 * all of them were opened.  By having a single journal transaction,
 * we eliminate duplicate reservations for common blocks like the
 * superblock and group descriptors or bitmaps.
 *
 * We will start the transaction here, but each prepare_write will
 * add a refcount to the transaction, and each commit_write will
 * remove a refcount.  The transaction will be closed when all of
 * the pages have been written.
 */
static void *fsfilt_ext3_brw_start(int objcount, struct fsfilt_objinfo *fso,
                                   int niocount, struct niobuf_local *nb,
                                   void *desc_private, int logs)
{
        journal_t *journal;
        handle_t *handle;
        int needed;
        ENTRY;

        LASSERT(current->journal_info == desc_private);
        journal = EXT3_SB(fso->fso_dentry->d_inode->i_sb)->s_journal;
        needed = fsfilt_ext3_credits_needed(objcount, fso, niocount, nb);

        /* The number of blocks we could _possibly_ dirty can very large.
         * We reduce our request if it is absurd (and we couldn't get that
         * many credits for a single handle anyways).
         *
         * At some point we have to limit the size of I/Os sent at one time,
         * increase the size of the journal, or we have to calculate the
         * actual journal requirements more carefully by checking all of
         * the blocks instead of being maximally pessimistic.  It remains to
         * be seen if this is a real problem or not.
         */
        if (needed > journal->j_max_transaction_buffers) {
                CERROR("want too many journal credits (%d) using %d instead\n",
                       needed, journal->j_max_transaction_buffers);
                needed = journal->j_max_transaction_buffers;
        }

        LASSERTF(needed > 0, "can't start %d credit transaction\n", needed);
        lock_24kernel();
        handle = fsfilt_ext3_journal_start(fso->fso_dentry->d_inode, needed);
        unlock_24kernel();
        if (IS_ERR(handle)) {
                CERROR("can't get handle for %d credits: rc = %ld\n", needed,
                       PTR_ERR(handle));
        } else {
                LASSERT(handle->h_buffer_credits >= needed);
                LASSERT(current->journal_info == handle);
        }

        RETURN(handle);
}

static int fsfilt_ext3_extend(struct inode *inode, unsigned int nblocks,void *h)
{
       handle_t *handle = h;

       /* fsfilt_extend called with nblocks = 0 for testing in special cases */
       if (nblocks == 0) {
               handle->h_buffer_credits = 0;
               CWARN("setting credits of handle %p to zero by request\n", h);
       }

       if (handle->h_buffer_credits > nblocks)
                return 0;
       if (journal_extend(handle, nblocks) == 0)
                return 0;

       ext3_mark_inode_dirty(handle, inode);
       return journal_restart(handle, nblocks);
}

static int fsfilt_ext3_commit(struct inode *inode, void *h, int force_sync)
{
        int rc;
        handle_t *handle = h;

        LASSERT(current->journal_info == handle);
        if (force_sync)
                handle->h_sync = 1; /* recovery likes this */

        lock_24kernel();
        rc = fsfilt_ext3_journal_stop(handle);
        unlock_24kernel();

        return rc;
}

static int fsfilt_ext3_commit_async(struct inode *inode, void *h,
                                    void **wait_handle)
{
        unsigned long tid;
        transaction_t *transaction;
#if LINUX_VERSION_CODE < KERNEL_VERSION(2,5,0)
        unsigned long rtid;
#endif
        handle_t *handle = h;
        journal_t *journal;
        int rc;

        LASSERT(current->journal_info == handle);

        lock_24kernel();
        transaction = handle->h_transaction;
        journal = transaction->t_journal;
        tid = transaction->t_tid;
        /* we don't want to be blocked */
        handle->h_sync = 0;
        rc = fsfilt_ext3_journal_stop(handle);
        if (rc) {
                CERROR("error while stopping transaction: %d\n", rc);
                unlock_24kernel();
                return rc;
        }
#if LINUX_VERSION_CODE < KERNEL_VERSION(2,5,0)
        rtid = log_start_commit(journal, transaction);
        if (rtid != tid)
                CERROR("strange race: %lu != %lu\n",
                       (unsigned long) tid, (unsigned long) rtid);
#else
        log_start_commit(journal, tid);
#endif
        unlock_24kernel();

        *wait_handle = (void *) tid;
        CDEBUG(D_INODE, "commit async: %lu\n", (unsigned long) tid);
        return 0;
}

static int fsfilt_ext3_commit_wait(struct inode *inode, void *h)
{
        journal_t *journal = EXT3_JOURNAL(inode);
        tid_t tid = (tid_t)(long)h;

        CDEBUG(D_INODE, "commit wait: %lu\n", (unsigned long) tid);
        if (unlikely(is_journal_aborted(journal)))
                return -EIO;

        log_wait_commit(EXT3_JOURNAL(inode), tid);

        if (unlikely(is_journal_aborted(journal)))
                return -EIO;
        return 0;
}

static int fsfilt_ext3_setattr(struct dentry *dentry, void *handle,
                               struct iattr *iattr, int do_trunc)
{
        struct inode *inode = dentry->d_inode;
        int rc = 0;

        lock_24kernel();

        /* Avoid marking the inode dirty on the superblock list unnecessarily.
         * We are already writing the inode to disk as part of this
         * transaction and want to avoid a lot of extra inode writeout
         * later on. b=9828 */
        if (iattr->ia_valid & ATTR_SIZE && !do_trunc) {
                /* ATTR_SIZE would invoke truncate: clear it */
                iattr->ia_valid &= ~ATTR_SIZE;
                EXT3_I(inode)->i_disksize = iattr->ia_size;
                i_size_write(inode, iattr->ia_size);

                if (iattr->ia_valid & ATTR_UID)
                        inode->i_uid = iattr->ia_uid;
                if (iattr->ia_valid & ATTR_GID)
                        inode->i_gid = iattr->ia_gid;
                if (iattr->ia_valid & ATTR_ATIME)
                        inode->i_atime = iattr->ia_atime;
                if (iattr->ia_valid & ATTR_MTIME)
                        inode->i_mtime = iattr->ia_mtime;
                if (iattr->ia_valid & ATTR_CTIME)
                        inode->i_ctime = iattr->ia_ctime;
                if (iattr->ia_valid & ATTR_MODE) {
                        inode->i_mode = iattr->ia_mode;

                        if (!in_group_p(inode->i_gid) && !capable(CAP_FSETID))
                                inode->i_mode &= ~S_ISGID;
                }

                inode->i_sb->s_op->dirty_inode(inode);

                goto out;
        }

        /* Don't allow setattr to change file type */
        if (iattr->ia_valid & ATTR_MODE)
                iattr->ia_mode = (inode->i_mode & S_IFMT) |
                                 (iattr->ia_mode & ~S_IFMT);

        /* We set these flags on the client, but have already checked perms
         * so don't confuse inode_change_ok. */
        iattr->ia_valid &= ~(ATTR_MTIME_SET | ATTR_ATIME_SET);

        if (inode->i_op->setattr) {
                rc = inode->i_op->setattr(dentry, iattr);
        } else {
                rc = inode_change_ok(inode, iattr);
                if (!rc)
                        rc = inode_setattr(inode, iattr);
        }

 out:
        unlock_24kernel();
        RETURN(rc);
}

static int fsfilt_ext3_iocontrol(struct inode * inode, struct file *file,
                                 unsigned int cmd, unsigned long arg)
{
        int rc = 0;
        ENTRY;

        /* FIXME: Can't do this because of nested transaction deadlock */
        if (cmd == EXT3_IOC_SETFLAGS && (*(int *)arg) & EXT3_JOURNAL_DATA_FL) {
                CERROR("can't set data journal flag on file\n");
                RETURN(-EPERM);
        }

        if (inode->i_fop->ioctl)
                rc = inode->i_fop->ioctl(inode, file, cmd, arg);
        else
                RETURN(-ENOTTY);

        RETURN(rc);
}

static int fsfilt_ext3_set_md(struct inode *inode, void *handle,
                              void *lmm, int lmm_size, const char *name)
{
        int rc;

        LASSERT(TRYLOCK_INODE_MUTEX(inode) == 0);

        lock_24kernel();
        rc = ext3_xattr_set_handle(handle, inode, EXT3_XATTR_INDEX_TRUSTED,
                                   name, lmm, lmm_size, 0);

        unlock_24kernel();

        if (rc && rc != -EROFS)
                CERROR("error adding MD data to inode %lu: rc = %d\n",
                       inode->i_ino, rc);
        return rc;
}

/* Must be called with i_mutex held */
static int fsfilt_ext3_get_md(struct inode *inode, void *lmm, int lmm_size,
                              const char *name)
{
        int rc;

        LASSERT(TRYLOCK_INODE_MUTEX(inode) == 0);
        lock_24kernel();

        rc = ext3_xattr_get(inode, EXT3_XATTR_INDEX_TRUSTED,
                            name, lmm, lmm_size);
        unlock_24kernel();

        /* This gives us the MD size */
        if (lmm == NULL)
                return (rc == -ENODATA) ? 0 : rc;

        if (rc < 0) {
                CDEBUG(D_INFO, "error getting EA %d/%s from inode %lu: rc %d\n",
                       EXT3_XATTR_INDEX_TRUSTED, name,
                       inode->i_ino, rc);
                memset(lmm, 0, lmm_size);
                return (rc == -ENODATA) ? 0 : rc;
        }

        return rc;
}

#if (LINUX_VERSION_CODE > KERNEL_VERSION(2,5,0))
static int fsfilt_ext3_send_bio(int rw, struct inode *inode, struct bio *bio)
{
        submit_bio(rw, bio);
        return 0;
}
#else
static int fsfilt_ext3_send_bio(int rw, struct inode *inode, struct kiobuf *bio)
{
        int rc, blk_per_page;

        rc = brw_kiovec(rw, 1, &bio, inode->i_dev,
                        KIOBUF_GET_BLOCKS(bio), 1 << inode->i_blkbits);
        /*
         * brw_kiovec() returns number of bytes actually written. If error
         * occurred after something was written, error code is returned though
         * kiobuf->errno. (See bug 6854.)
         */

        blk_per_page = CFS_PAGE_SIZE >> inode->i_blkbits;

        if (rc != (1 << inode->i_blkbits) * bio->nr_pages * blk_per_page) {
                CERROR("short write?  expected %d, wrote %d (%d)\n",
                       (1 << inode->i_blkbits) * bio->nr_pages * blk_per_page,
                       rc, bio->errno);
        }
        if (bio->errno != 0) {
                CERROR("IO error. Wrote %d of %d (%d)\n",
                       rc,
                       (1 << inode->i_blkbits) * bio->nr_pages * blk_per_page,
                       bio->errno);
                rc = bio->errno;
        }

        return rc;
}
#endif

static ssize_t fsfilt_ext3_readpage(struct file *file, char *buf, size_t count,
                                    loff_t *off)
{
        struct inode *inode = file->f_dentry->d_inode;
        int rc = 0;

        if (S_ISREG(inode->i_mode))
                rc = file->f_op->read(file, buf, count, off);
        else {
                const int blkbits = inode->i_sb->s_blocksize_bits;
                const int blksize = inode->i_sb->s_blocksize;

                CDEBUG(D_EXT2, "reading "LPSZ" at dir %lu+%llu\n",
                       count, inode->i_ino, *off);
                while (count > 0) {
                        struct buffer_head *bh;

                        bh = NULL;
                        if (*off < i_size_read(inode)) {
                                int err = 0;

                                bh = ext3_bread(NULL, inode, *off >> blkbits,
                                                0, &err);

                                CDEBUG(D_EXT2, "read %u@%llu\n", blksize, *off);

                                if (bh) {
                                        memcpy(buf, bh->b_data, blksize);
                                        brelse(bh);
                                } else if (err) {
                                        /* XXX in theory we should just fake
                                         * this buffer and continue like ext3,
                                         * especially if this is a partial read
                                         */
                                        CERROR("error read dir %lu+%llu: %d\n",
                                               inode->i_ino, *off, err);
                                        RETURN(err);
                                }
                        }
                        if (!bh) {
                                struct ext3_dir_entry_2 *fake = (void *)buf;

                                CDEBUG(D_EXT2, "fake %u@%llu\n", blksize, *off);
                                memset(fake, 0, sizeof(*fake));
                                fake->rec_len = cpu_to_le16(blksize);
                        }
                        count -= blksize;
                        buf += blksize;
                        *off += blksize;
                        rc += blksize;
                }
        }

        return rc;
}

static void fsfilt_ext3_cb_func(struct journal_callback *jcb, int error)