Skip to content

Commit

Permalink
Bug fix: [zfs-linux/zfs GH-57] Writing with buffered IO and and readi…
Browse files Browse the repository at this point in the history 
…ng with mmap corrupts the data.
  • Loading branch information
Prasad Joshi committed Dec 19, 2010
1 parent 5be4800 commit ee3f1d3
Show file tree
Hide file tree
Showing 2 changed files with 47 additions and 68 deletions.
5 changes: 3 additions & 2 deletions module/lzfs_super.c
View file
Expand Up @@ -61,9 +61,10 @@ extern void lzfs_zfsctl_destroy(vfs_t *);

static void lzfs_delete_vnode(struct inode *inode)
{
truncate_inode_pages(&inode->i_data, 0);
loff_t oldsize = i_size_read(inode);
i_size_write(inode, 0);
truncate_pagecache(inode, oldsize, 0);
clear_inode(inode);
inode->i_size = 0;
}

static void
Expand Down
110 changes: 44 additions & 66 deletions module/lzfs_vnops.c
View file
Expand Up @@ -798,6 +798,10 @@ lzfs_vnop_read (struct file *filep, char __user *buf, size_t len, loff_t *ppos)
}
}
nr = nr - offset;

if (mapping_writably_mapped(mapping))
flush_dcache_page(page);

mark_page_accessed(page);
prev_index = index;

Expand Down Expand Up @@ -916,54 +920,48 @@ ssize_t
lzfs_vnop_write (struct file *filep, const char __user *buf, size_t len,
loff_t *ppos)
{
int err;
vnode_t *vp = NULL;
ssize_t i_size;
ssize_t rc;

/* PAGE CACHE SUPPORT VARIABLES */
struct address_space *mapping = filep->f_mapping;
struct inode *inode = mapping->host;
pgoff_t index;
unsigned long written = 0;
unsigned long written;
unsigned long offset;

const char *user_buf = buf;

/* pos_append would be equal to ppos, but if file is opened
* with O_APPEND flag pos_append would be assigned size of inode */
loff_t pos_append = 0;
loff_t pos_append;

SENTRY;

vp = LZFS_ITOV(inode);

if (likely(!(vp->v_flag & VMMAPPED))) {
/* file is not memory mmapped, pass write directly to ZFS */
ssize_t rc;
rc = lzfs_write(vp, filep->f_flags, buf, len, *ppos, UIO_USERSPACE);
if (likely(rc > 0))
*ppos += rc;
rc = lzfs_write(vp, filep->f_flags, buf, len, *ppos, UIO_USERSPACE);
if (unlikely(rc < 0)) {
tsd_exit();
SEXIT;
return rc;
}

i_size = i_size_read(inode);

pos_append = *ppos;
if (filep->f_flags & FAPPEND) {
/* file is opened with the O_APPEND flag */
pos_append = i_size;
*ppos += rc;

if (likely(!(vp->v_flag & VMMAPPED))) {
/* file is not memory mmapped, pass write directly to ZFS */
tsd_exit();
SEXIT;
return rc;
}

index = pos_append >> PAGE_CACHE_SHIFT;
offset = pos_append & ~PAGE_CACHE_MASK;
written = 0;

while (1) {
struct page *page = NULL;
char *page_buf;
ssize_t size;

ssize_t tmp;

if (written >= len)
break;
Expand All @@ -976,78 +974,61 @@ lzfs_vnop_write (struct file *filep, const char __user *buf, size_t len,
page = find_lock_page(mapping, index);
if (likely(!page)) {
#if 0
printk ("%s: Page not cached. ppos=%llu, pos_append=%llu, offset=%ld, len=%ld, size=%ld\n",
__FUNCTION__, *ppos, pos_append, offset, len, size);
printk ("%s: Page not cached. ppos=%llu, pos_append=%llu, offset=%ld, "
"len=%ld, size=%ld\n", __FUNCTION__, *ppos, pos_append, offset,
len, size);
#endif
goto no_cached_page;
}
#if 0
printk ("%s: Page cached. ppos=%llu, pos_append=%llu, offset=%ld, len=%ld, size=%ld\n",
__FUNCTION__, *ppos, pos_append, offset, len, size);
printk ("%s: Page cached. ppos=%llu, pos_append=%llu, offset=%ld, len=%ld, "
"size=%ld\n", __FUNCTION__, *ppos, pos_append, offset, len, size);
#endif

if (mapping_writably_mapped(mapping))
flush_dcache_page(page);

pagefault_disable();

/* copy the data */
BUG_ON(!in_atomic());
page_buf = kmap_atomic(page, KM_USER0);
if (copy_from_user(page_buf+offset, user_buf, size)) {
kunmap(page);
unlock_page(page);
err = -EFAULT;
goto out_error;
}
tmp = __copy_from_user_inatomic(page_buf + offset, user_buf, size);
kunmap_atomic(page_buf, KM_USER0);
pagefault_enable();

flush_dcache_page(page);
if (tmp) {
unlock_page(page);
tsd_exit();
SEXIT;
return -EFAULT;
}

mark_page_accessed(page);

/* do i need to modify the inode size */
SetPageUptodate(page);
ClearPageError(page);
// set_page_dirty(page);
/* data is written to disk and then a page is modified
i.e. page is not dirty */
page_clear_dirty(page);
unlock_page(page);
page_cache_release(page);
zfs_file_modified(vp);
// zfs_file_modified(vp);

balance_dirty_pages_ratelimited(mapping);
// balance_dirty_pages_ratelimited(mapping);

no_cached_page:
user_buf += size;
written += size;
*ppos += size;
pos_append += size;
index = pos_append >> PAGE_CACHE_SHIFT;
if (*ppos != pos_append) {
/* file is opened with the O_APPEND flag */
BUG_ON(!(filep->f_flags & FAPPEND));
i_size_write(inode, i_size + size);
}
offset = 0;
continue;

no_cached_page:
size = lzfs_write(vp, filep->f_flags, user_buf, size, *ppos, UIO_USERSPACE);
if (unlikely(size < 0)) {
err = size;
goto out_error;
}

*ppos += size;
pos_append += size;
user_buf += size;
written += size;
index = *ppos >> PAGE_CACHE_SHIFT;
offset = 0;
}

tsd_exit();
SEXIT;
return ((ssize_t) written);
out_error:
tsd_exit();
SEXIT;
return err;
return rc;
}

/*
Expand Down Expand Up @@ -1087,7 +1068,9 @@ int lzfs_file_mmap(struct file * file, struct vm_area_struct * vma)
if (rc < 0)
return rc;

mutex_enter(&vp->v_lock);
vp->v_flag |= VMMAPPED;
mutex_exit(&vp->v_lock);
return rc;
}

Expand Down Expand Up @@ -1262,8 +1245,6 @@ static int lzfs_readpage(struct file *file, struct page *page)
vnode_t *vp = NULL;
unsigned long fillsize = 0;
ssize_t rc;
struct iovec iov;
uio_t uio;

BUG_ON(!PageLocked(page));
// printk(KERN_ERR "%s: ==>\n", __FUNCTION__);
Expand All @@ -1286,9 +1267,6 @@ static int lzfs_readpage(struct file *file, struct page *page)
goto out;
}

bzero(&iov, sizeof(struct iovec));
bzero(&uio, sizeof(uio_t));

if (offset < i_size) {
i_size -= offset;
fillsize = i_size > PAGE_CACHE_SIZE ? PAGE_CACHE_SIZE : i_size;
Expand Down Expand Up @@ -1379,7 +1357,7 @@ static int lzfs_writepage(struct page *page, struct writeback_control *wbc)
}
kunmap(page);
out:
// page_clear_dirty(page);
page_clear_dirty(page);
unlock_page(page);
// printk(KERN_ERR "%s: <==\n", __FUNCTION__);
return err;
Expand Down

0 comments on commit ee3f1d3

Please sign in to comment.