| Message ID | 20151117201614.15053.62376.stgit@dwillia2-desk3.jf.intel.com (mailing list archive) |
|---|---|
| State | Superseded, archived |
| Headers | show |
On Tue 17-11-15 12:16:14, Dan Williams wrote: > Currently dax mappings survive block_device shutdown. While page cache > pages are permitted to be read/written after the block_device is torn > down this is not acceptable in the dax case as all media access must end > when the device is disabled. The pfn backing a dax mapping is permitted > to be invalidated after bdev shutdown and this is indeed the case with > brd. > > When a dax capable block_device driver calls del_gendisk() in its > shutdown path, or a filesystem evicts an inode it needs to ensure that > all the pfns that had been mapped via bdev_direct_access() are unmapped. > This is different than the pagecache backed case where > truncate_inode_pages() is sufficient to end I/O to pages mapped to a > dying inode. > > Since dax bypasses the page cache we need to unmap in addition to > truncating pages. Also, since dax mappings are not accounted in the > mapping radix we uncoditionally truncate all inodes with the S_DAX flag. > Likely when we add support for dynamic dax enable/disable control we'll > have infrastructure to detect if the inode is unmapped and can skip the > truncate. > > Cc: <stable@vger.kernel.org> > Cc: Jan Kara <jack@suse.com> > Cc: Dave Chinner <david@fromorbit.com> > Cc: Matthew Wilcox <willy@linux.intel.com> > Cc: Ross Zwisler <ross.zwisler@linux.intel.com> > Signed-off-by: Dan Williams <dan.j.williams@intel.com> > --- > fs/inode.c | 27 +++++++++++++++++++++++++++ > mm/truncate.c | 13 +++++++++++-- > 2 files changed, 38 insertions(+), 2 deletions(-) ... > @@ -433,7 +434,15 @@ void truncate_inode_pages_final(struct address_space *mapping) > spin_lock_irq(&mapping->tree_lock); > spin_unlock_irq(&mapping->tree_lock); > > - truncate_inode_pages(mapping, 0); > + /* > + * In the case of DAX we also need to unmap the inode > + * since the pfn backing the mapping may be invalidated > + * after this returns > + */ > + if (IS_DAX(inode)) > + truncate_pagecache(inode, 0); > + else > + truncate_inode_pages(mapping, 0); > } Hum, I don't get this. truncate_inode_pages_final() gets called when inode has no more users. So there are no mappings of the inode. So how could truncate_pagecache() possibly make a difference? Honza
diff --git a/fs/inode.c b/fs/inode.c index 1be5f9003eb3..1029e033e991 100644 --- a/fs/inode.c +++ b/fs/inode.c @@ -579,6 +579,18 @@ static void dispose_list(struct list_head *head) } } +static void truncate_list(struct list_head *head) +{ + struct inode *inode, *_i; + + list_for_each_entry_safe(inode, _i, head, i_lru) { + list_del_init(&inode->i_lru); + truncate_pagecache(inode, 0); + iput(inode); + cond_resched(); + } +} + /** * evict_inodes - evict all evictable inodes for a superblock * @sb: superblock to operate on @@ -642,6 +654,7 @@ int invalidate_inodes(struct super_block *sb, bool kill_dirty) int busy = 0; struct inode *inode, *next; LIST_HEAD(dispose); + LIST_HEAD(truncate); spin_lock(&sb->s_inode_list_lock); list_for_each_entry_safe(inode, next, &sb->s_inodes, i_sb_list) { @@ -655,6 +668,19 @@ int invalidate_inodes(struct super_block *sb, bool kill_dirty) busy = 1; continue; } + if (IS_DAX(inode) && atomic_read(&inode->i_count)) { + /* + * dax mappings can't live past this invalidation event + * as there is no page cache present to allow the data + * to remain accessiable. + */ + __iget(inode); + inode_lru_list_del(inode); + spin_unlock(&inode->i_lock); + list_add(&inode->i_lru, &truncate); + busy = 1; + continue; + } if (atomic_read(&inode->i_count)) { spin_unlock(&inode->i_lock); busy = 1; @@ -669,6 +695,7 @@ int invalidate_inodes(struct super_block *sb, bool kill_dirty) spin_unlock(&sb->s_inode_list_lock); dispose_list(&dispose); + truncate_list(&truncate); return busy; } diff --git a/mm/truncate.c b/mm/truncate.c index 76e35ad97102..ff1fb3b0980e 100644 --- a/mm/truncate.c +++ b/mm/truncate.c @@ -402,6 +402,7 @@ EXPORT_SYMBOL(truncate_inode_pages); */ void truncate_inode_pages_final(struct address_space *mapping) { + struct inode *inode = mapping->host; unsigned long nrshadows; unsigned long nrpages; @@ -423,7 +424,7 @@ void truncate_inode_pages_final(struct address_space *mapping) smp_rmb(); nrshadows = mapping->nrshadows; - if (nrpages || nrshadows) { + if (nrpages || nrshadows || IS_DAX(inode)) { /* * As truncation uses a lockless tree lookup, cycle * the tree lock to make sure any ongoing tree @@ -433,7 +434,15 @@ void truncate_inode_pages_final(struct address_space *mapping) spin_lock_irq(&mapping->tree_lock); spin_unlock_irq(&mapping->tree_lock); - truncate_inode_pages(mapping, 0); + /* + * In the case of DAX we also need to unmap the inode + * since the pfn backing the mapping may be invalidated + * after this returns + */ + if (IS_DAX(inode)) + truncate_pagecache(inode, 0); + else + truncate_inode_pages(mapping, 0); } } EXPORT_SYMBOL(truncate_inode_pages_final);
Currently dax mappings survive block_device shutdown. While page cache pages are permitted to be read/written after the block_device is torn down this is not acceptable in the dax case as all media access must end when the device is disabled. The pfn backing a dax mapping is permitted to be invalidated after bdev shutdown and this is indeed the case with brd. When a dax capable block_device driver calls del_gendisk() in its shutdown path, or a filesystem evicts an inode it needs to ensure that all the pfns that had been mapped via bdev_direct_access() are unmapped. This is different than the pagecache backed case where truncate_inode_pages() is sufficient to end I/O to pages mapped to a dying inode. Since dax bypasses the page cache we need to unmap in addition to truncating pages. Also, since dax mappings are not accounted in the mapping radix we uncoditionally truncate all inodes with the S_DAX flag. Likely when we add support for dynamic dax enable/disable control we'll have infrastructure to detect if the inode is unmapped and can skip the truncate. Cc: <stable@vger.kernel.org> Cc: Jan Kara <jack@suse.com> Cc: Dave Chinner <david@fromorbit.com> Cc: Matthew Wilcox <willy@linux.intel.com> Cc: Ross Zwisler <ross.zwisler@linux.intel.com> Signed-off-by: Dan Williams <dan.j.williams@intel.com> --- fs/inode.c | 27 +++++++++++++++++++++++++++ mm/truncate.c | 13 +++++++++++-- 2 files changed, 38 insertions(+), 2 deletions(-) -- To unsubscribe from this list: send the line "unsubscribe linux-block" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html